Open access peer-reviewed chapter

Two-Phase Network Data Envelopment Analysis: An Example of Bank Performance Assessment

By Yu-cheng Liu

Submitted: October 23rd 2017Reviewed: February 5th 2018Published: November 5th 2018

DOI: 10.5772/intechopen.74933

Downloaded: 713


Data envelopment analysis (DEA) models assess decision-making units (DMUs), which directly convert multiple inputs into multiple outputs. Network DEA models have been studied extensively. However, the performance indices that link the two stages are assumed to be fixed or non-discretionary; their values are not adjustable. These models only assumed that the reductions on the inputs and additions on the outputs would improve the overall efficiency. But in the real world, the link is always adjustable. “Free links” means that the intermediate items are adjustable or discretionary, and each DMU can be increased or decreased from the observed one. The current chapter introduces a two-phase procedure with free links to assess system performance, Phase-I is a proposed slack-based measurement (SBM) model to partition the links into two sets: as-input and as-output. Phase-II is a modified SBM model to determine the slack of each input, as-input link, output and as-output link. This proposed model counts the slacks associated with the intermediate items in the efficiency scores and determines the entire system performance by the directional distance function. It is validated using network procedure and assesses the performance of supply chain management system.


  • data envelopment analysis
  • performance measure
  • directional distance function
  • network DEA
  • slack-based measure

1. Introduction

The data envelopment analysis (DEA) models assess a set of homogeneous decision-making units (DMUs) that convert inputs into outputs. Fewer input values and more output values are desired and DMUs may be classified as being either efficient or inefficient. Tone and Tsutsui [1, 2] introduce network and dynamic DEA and categorize links into two types—“fixed links” and “free links”. The free links mean the links are adjustable; each DMU can be increased or decreased from the observed one and identifies the improvement target of each inefficient DMU on the frontier that is constructed by the efficient DMUs.

Seiford and Zhu [3] and Zhu [4] have introduced a two-stage process to measure the profitability and marketability of 55 US commercial banks and top Fortune 500 companies, respectively. They propose the effect of bank size on profitability and marketability through evaluating both technical and scale efficiencies. Sexton and Lewis [5] use a two-stage approach to evaluate the scores of American Major League Baseball teams. There are many other cases in which the whole operation is separated into more than two processes. These may have a series structure, a parallel structure, or a mixture of these. These structures are generally called network structures and the DEA technique to measure the efficiency of systems with a network structure is called network DEA (Färe & Grosskopf [6]). Färe and Whittaker [7] and Färe and Grosskopf [8] introduce models to compute the efficiency scores of sub-processes in network-structured DEA problems. Lewis and Sexton [9] introduce a network DEA model which focuses efficiency-enhancing strategies on individual stages of the production process. Kao and Hwang [10] introduce a framework for breaking down the efficiency of the entire process into the product of the efficiencies of the two-stage process. It assumes that the weights on the links are the same for the two stages, that is, the weights on the outputs in the first stage are assumed to be equal to the weights on the inputs in the second stage. In the real world, the relative weight of each stage is determined corresponding to its importance. Thus, the different weights in the entire system are mentioned in recent studies. Chen et al. [11] mentions that the overall efficiency scores resulting from Kao and Hwang [10] are not direct indicators of potential input reductions or output increases not realized by the inefficient DMUs. They develop an approach to determine the DEA frontier or DEA projections for inefficient DMUs. Chen et al. [12] note that the envelopment-based network DEA model should be used for determining the frontier projection for inefficient DMUs, whereas the multiplier-based network DEA model should be used for determining the divisional efficiency because it does not account for the intermediate links. Kao [13] proposes a dynamic DEA model to measure system and period efficiencies at the same time for multi-period systems. Chang et al. [14] take into account the ownership structure of networks in constructing effective network DEA models and accordingly develop three ownership-specified (centralized, distributed, and hybrid) network DEA models. Huang et al. [15] proposed a two-stage network model with bad outputs and supper efficiency (US-NSBM). Empirical comparisons show that the US-NSBM may be promising and practical for taking the nonperforming loans into account and being able to rank all samples.

However, these approaches do not count the slacks associated with the intermediate items in the efficiency scores. Consequently, the efficiency scores are greater than the actual efficiency. In addition, there is no DMU with an efficiency score equal to 1 because the properties of intermediate performance evaluation items would lead to conflicts. For instance, in the two-stage process problem, Stage-2 may have to reduce inputs (links) to achieve an efficient status. However, doing so would lead to a reduction in outputs in Stage-1, thereby reducing the efficiency of Stage-1. In other words, there are still two efficiency frontiers for the two sub-processes. One may desire a single frontier for the entire production system.

“Link” cannot be adjusted freely in a radial model which adjusts the inputs and outputs by the efficiency scores in a two-stage process. For this model, the entire system efficiency cannot be improved by adjusting links, see Kao and Hwang [10] and Lewis and Sexton [9]. “Link” that applies in a non-radial model has been discussed in recent years. Tone and Tsutsui [1] introduce a network DEA and categorize links into two types—“fixed links” and “free links.” “Free links” means the intermediate items are adjustable or discretionary; each DMU can be increased or decreased from the observed one and is free to assign each individual link to one of the three characteristics: as-input, as-output, or non-discretionary so that the entire system efficiency could be maximized. “Fixed links” means the intermediate products are beyond the control of DMUs. In the radial model, “links” cannot be adjusted freely, which adjust the inputs and outputs by the efficiency scores in a two-stage process. Tone and Tsutsui [2] introduce the dynamic slack-based measure (DSBM) model and the incorporation of slacks with freeand fixedlinks into the efficiency score. They categorize the links into four types: desirable, undesirable, discretionary (free), and non-discretionary (fixed). The article incorporates the slacks of free links into the efficiency score in two ways: an ex-post approach (adjusted score) and incorporation through 0–1 MIP. The ex-post approach includes a two-phase procedure. Tone and Tsutsui [1] introduce the links are discretionary regarding their status, as-input or as-output. Liu and Liu [16, 17] adopt VGM and GBM models to assess the performance of supply chain management.

Chambers et al. [18] introduced the directional distance function (DDF) based on the Luenberger benefit function to obtain the technical efficiency by increasing the outputs and reducing the inputs simultaneously. Later, Chambers et al. [19] introduced the DDF of DEA to measure the technical efficiency. This chapter develops a model for an improved efficiency measure through directional distance formulation of data envelopment analysis.

The contribution and innovative progress for this chapter are (1) creating a new SBM model and converting multi-efficiency frontiers for the separation processes to an aggregation efficiency frontier for the entire production system and (2) adopting free links application and introducing DDF with a virtual gap diagram to assess the performance of the entire system. The rest of this chapter is organized as follows. The proposed two-phase two-stage performance evaluation models and DDF are presented in Section 2. Because the uniqueness of the optimal solution is important, we report an experiment on this subject using a real-world bank performance assessment in Section 3. We conclude this chapter in the last section.


2. The proposed two-phase two-stage performance evaluation

Jdenotes the set of homogeneous decision-making units of a network process that are evaluated by a set of inputs, I, a set of free links, Dfree, and a set of outputs R. DMUorepresents the DMU under evaluation. To maximize the system efficiency score of DMUo, each link in set Dfreeis “free” to be assigned to one of the subsets – Do, Dofree, and Do+if it is as-input, free link, and as-output, respectively. Figure 1 depicts the two-phase procedure to evaluate the performance of DMUs using the two-stage and network processes. This two-phase procedure contains two slack-based DEA linear programming models. Phase-I sets all links in set Dfreeto discretionary and the objective is to determine the maximum slack values on each input and output so that the weights of each DMU in each stage can be assigned. The set Dfreeis partitioned into two subsets Do, Dofreeand Do+. The output of Phase-I will indicate that several links in set Dofreeshould be assigned to sets Doand Do+. The target of Phase-II is thus to determine the maximum reduction value on each input and link in sets Iand Doand the addition value on each output and link in sets Rand Do+such that the weights of each DMU in the system can be assigned. The target of each input, free link, and output on the frontier is identified.

Figure 1.

The flow of two-phase procedure.

2.1. Two stages: Phase-I

Envelopment via the SBM fractional programming model [M1] is used to measure the relative performance of DMUo. The decision variables of slack are the values to be subtracted at the ith input and the value to be added at the rth output, respectively. The decision variables denote the weights of DMUjat Stage-1 and Stage-2, respectively. The right-hand side of constraints (1.2) and (1.3) expresses the targets of the inputs and links (outputs) at Stage-1. Each link din set Dfreeis free to increase the slack szdfree+or decrease the slack szdfree. These slacks are non-discretionary and not counted in the objective function. The decision variables on the left-hand side are λ1j,jJ. Similarly, the right-hand side of constraints (1.4) and (1.5) expresses the targets of the links (inputs) and outputs at Stage-2. The decision variables on the left-hand side are λ2j,jJ. Constraints (1.3) and (1.4) indicate that their right-hand sides are equal; each link din set Dfreecan be freely adjusted to reach its single target at Stage-1 and Stage-2 simultaneously. The optimal slack value of link din set Dfree, szdfree+szdfree+, is >0, <0 or =0, in which case link dof DMUois assigned to sets Do+, Do, and Dofree, respectively. Two tasks remain for DMUo. The first task is to assign all elements in the set to either Do+or Do. The second task is to place the slacks of link din sets and add them to its aggregate performance score. Phase-II of our solving procedure addresses the first task.



For the two-phase procedure which is depicted in Figure 1, Phase-I is to determine the maximum slack values on each input and output; [M1] presents this purpose and adopts Eq. (1.3) and (1.4) to distinguish links to be as-input, discretionary, and as-output, which express as three subsets, Do, Dofreeand Do+, respectively. The aim of Phase-I is to assign each element in set Dofreeto either Do+or Do. [M2] is repeated until set Dofreebecomes empty. The fractional programming model [M2] measures the overall efficiency ρoII. The decision variables πj1and πj2denote the weights in Stage-1 and Stage-2, respectively, of DMUjin evaluating DMUo.



The solution of [M2] for each link din set Dofreeis one of the following cases, szdfree+szdfree+<0, szdfree+szdfree+>0, and szdfree+szdfree+=0. Next, dis assigned to set Do, Do+, and Dofreeaccordingly.

2.2. Two stages: Phase-II

The results of Phase-I indicate that DMUoalready partitioned set Dofreeinto Doand Do+. The fractional programming model [M3] is an SBM model (Tone and Tsutsui [21]) that measures the efficiency of converting the sets of input and as-input indices IDointo the sets of output and as-output indices RDo+. The frontiers of πj1andπj2are converted into an entire system frontier πjin this phase.



The dual form of [M3] is expressed as [M4]. The decision variables of [M4] possess properties

, representing the weight assigned to the ith input and the rth output, respectively. The terms wd+and wdrepresent the weight assigned to link din sets Do+and Do, respectively.


vi,urfree insign,iI,rR;E29
wdfree insign,dDo;wd+free insign,dDo+.E30

Inequality (4.2) may be revised such that rRuryrj+dDo+wd+zdj/iIvixij+dDowdzdj1,and the constraint ensures that the maximum performance value of each DMUjis not greater than 1.

2.3. Proposed directional distance function approach

The directional distance function (DDF) measures the distance from a certain operation point (e.g., DMUo) to the efficient frontier of the technology along the positive semi-ray defined by vector g. Given a directional vector g=gXgY+, gX+I+Dand gY++R+D+. The objective function (4.1) can be modeled by using the DDF. We denote virtual input by (X= iIvixio+dDowdzdo) and virtual output by (Y= dDo+wd+zdo+rRuryro) which are identified by specifying a directional vector g. The objective function (4.1) can be converted to (4.9) which is to minimize the virtual input and maximize the virtual output to reach the efficient frontier.


The graph technology can be represented by T=XYX+I+DoY+R+Do+. The optimal solution of virtual gap Δoexpresses as DDF: DgXYgXgY+=supiIxiovi×gX+dDozdowd×gXrRyrour×gY++dDo+zdowd+×gY+TXY. This chapter defines a virtual gap diagram; the summation of input and as-input is the x-axis (iIvixio+dDowdzdo) and the summation of output and as-output is the y-axis (dDo+wd+zdo+rRuryro). The geometry on the virtual gap diagram is the slope of the line from DMUoto origin. To evaluate different DMUo, one may directly compare their vectors of weights; virtual gap, Δo; virtual input and virtual as-input, ΔoI; and virtual as-output and virtual output, ΔoO. It is obvious that the minimum virtual gap “

” is equivalent to the maximum efficiency score of the entire network. It ensures that the nearest improvement target is found. Figure 2 depicts the virtual gap diagram; x-axis denotes the virtual input and y-axis denotes the virtual output.

Figure 2.

Virtual gap diagram.

2.4. Overall stage efficiencies

Similar to the SBM non-oriented models of Tone and Tsutsui [20], the solutions of Phase-II provide a reference set of DMUs for DMUo. The target for the performance items in sets I,Do+, Do, and Rcan be obtained using [E1]. The measured performance value Eois the best practice for DMUoin the overall two-stage process which is expressed as [E2].


These points are the projection of DMUoon the frontier.


The results of Phase-II, si,szd,szd+andsr+, are used to compute the efficiencies of Stage-1, E1, and Stage-2; E2is shown in the following two Eqs. (E3 and E4). For the efficiencies of Stage-1, the numerator is the summation of inputs (xio,iI) and as-input (zdo,dDo). The denominator is the as-output items (zdo,dDo+). Likewise, for the efficiencies of Stage-2, the numerator is the as-input item (zdo,dDo). The denominator is the summation of the as-output (zdo,dDo+) and outputs (yro,rR).


If set

is empty, the denominator is equal to 1.


If set

is empty, the numerator is equal to 1.

From Eqs. [E3] and [E4], we obtain the performance scores of Stage-1 and Stage-2, respectively, which identify the performance of each stage.

2.5. To extend two-stage to network process

Liu and Liu [16] extend the two-stage to network process. The network contains a set of sub-processes (nodes), H. The nodes are assigned ordinal numbers 1, 2, 3,…, n. Let A denote the set of network links. There are nhomogeneous DMUs in set J, named DMU1, DMU2,…, and DMUn, which are randomly processed by the sub-processes in set H. The network structure is depicted in Figure 3.

Figure 3.

Network structure.

2.5.1. Inputs and outputs

At each sub-process h, there is a set of input measures Ihthat flow into the network and a set of output measures Rhthat flow out of the network. For DMUjin set J, let xijh+Ihand yrjh+Rhdenote the volumes of the ith input measure and the rth output measure at the sub-process h, respectively. Let sihand srh+be the slack of the ith input and the rth output at sub-process h, respectively.

2.5.2. Links

Each sub-process may have links to other sub-processes. Let (h, k) denote the link between sub-processes hand k, h > k. Let Dhkdenote the set of link measures on link (h, k).zdjhk+Dhk++Dhkdenotes the volume of the dth link in set Dhk. Each DMU alternatively acts as the DMUothat is under evaluation. The volume of link don link (h, k) could be increased or decreased with a slack to improve the efficiency of DMUoas well.


3. Illustrative examples

This study adopts a dataset covering 24 non-life insurance companies in Taiwan from Kao and Hwang [10] to illustrate the proposed two-phase procedure. Table 1 summarizes the performance datasheet of 24 non-life insurance companies in Taiwan.

The inputs of the system:

Operation expenses (x1): salaries of the employees and various types of costs incurred in daily operation and.

Insurance expenses (x2): expenses paid to agencies, brokers, and solicitors and other expenses associated with marketing the service of insurance.

The links of the system:

Direct written premiums (z1): premiums received from insured clients.

Reinsurance premiums (z2): premiums received from ceding companies.

The outputs of the system:

Under-writing profit (y1): profit earned from the insurance business.

Investment profit (y2): profit earned from the investment portfolio.

3.1. Phase-I

Table 2 summarizes the results of Phase-I and Phase-II. In the Phase-I column, for example, when DMU1is being evaluated, DMUo = DMU1and the optimal solution of [M1] is sz1+=0, sz2+=549,067, sz1=877,494, and sz2=0. Therefore, Do+= {2}, Do= {1}, and Dofree= {}. The first row includes 2(549,067) and 1(877,494).

BanksDMUjOperation expensesInsurance expensesDirect written premiumsReinsurance premiumsUnderwriting profitInvestment profit
Taiwan Fire11,178,744673,5127,451,757856,735984,143681,687
Chung Kuo21,381,8221,352,75510,020,2741,812,8941,228,502834,754
Tai Ping31,177,494592,7904,776,548560,244293,613658,428
China Mariners4601,320594,2593,174,851371,863248,709177,331
Ming Tai83,789,0011,873,53017,267,2661,134,6003,899,530622,868
The First101,303,2491,298,4708,210,389504,5281,697,941554,806
South China141,396,002988,8887,396,396465,5091,401,200332,283
Cathay Century152,184,944651,06310,422,297749,8933,355,197555,482
Allianz President161,211,716415,0715,606,013402,881854,054197,947
North America19159,422182,3381,141,951483,291519,12146,857
Royal & Sunalliance2184,17126,224225,88840,54251,9506491
Mitsui Sumitomo24163,297235,094476,419644,816142,37016,976

Table 1.

Performance of 24 non-life insurance companies in Taiwan.

Table 2.

Phase-I solutions.

2*(549,067)+:*, the name of the link; +, the slack of the item.

When DMU4is being evaluated, DMUo = DMU4and the optimal solution of [M1] is sz1+=0, sz2+=516,873, sz1=0, and sz2=0. Therefore, Do+= {2}, Do= {}, and Dofree={1}. The fourth row includes 2(516,873) and 1(0) in the Phase-I column. The solution of Phase-I indicates that Do+= {2}, Do= {}, and Dofree={1}; the optimal solutions of [M4] are sz1free= 3,174,850 and sz1free+=0. This calculation indicates that the natural link, d = 1, acts as an “as-input” item and may have a better solution. Therefore, 1(3,174,850) is recorded under the

column of Phase-I. DMU7and DMU18have solution processes that are similar to DMU4.

3.2. Phase-II

Because each link may be “as-input” or “as-output”, the two links may have four possible combinations of Do+and Do. Table 4 shows the four categories A, B, C, and D and their link settings. As indicated in the first column of Table 3, 15, 3, 3, and 3 DMUs belong to Categories A, B, C, and D, respectively. For instance, DMU4 in Category A treats the first links as “as-input” (slack = 1,072,937) and is an undesirable output with respect to Stage-1 and a desirable input with respect to Stage-2. Meanwhile, the second set of links is “as-output” (slack = 135,818) and represents a desirable output with respect to Stage-1 but an undesirable input with respect to Stage-2.

Table 3.

Sets Do+and Doof each DMU in Phase-II.

Proceeding to Phase-II, which employs [M5], the optimal solutions for the evaluated DMU are listed in Table 4. The second column presents four efficiency scores obtained from (M5), Eqs. [E3], [E4], and [E2], which identify the Stage-1 efficiency (Eo1), Stage-2 efficiency (Eo2), and overall efficiency (Eo), respectively. The third column presents the reference DMUjof each evaluated DMUo. The projection points of DMUoon the frontier which obtain from Eopresents at right sides.

DMUjEo1Eo2EoReference DMUProjected Points

Table 4.

Final solutions’ summary.

The nine efficient DMUs, 1, 2, 3, 5, 11, 12, 20, 22, and 23, are consistent, with all of their performance scores in Stage-1 and Stage-2 being equal to one. The efficiency scores for the inefficient DMUs for both Stage-1 and Stage-2 are less than 1. For instance, DMU4has scores of 0.565 and 0.144 in Stage-1 and Stage-2. An obvious means of improving overall efficiency is to focus on Stage-2.

The virtual weight is expressed as vixij/VirtualInput,wdzdj/VirtualInputwd+zdj/VirtualOutputand uryrj/VirtualOutputwhere VirtualInput=iIvixij+dDowdzdjand VirtualOutput=rRuryrj+dDo+wd+zdj. These equations represent the percentage of each input or “as-input” link in the overall virtual input weight and the percentage of each output or “as-output” link in the overall virtual output weight. As shown at Table 5, it indicates the improving ratio of each inefficient DMU. For instance, DMU4has scores of 0.565 and 0.144 in Stage-1 and Stage-2. An obvious means of improving overall efficiency is to focus on Stage-2 output item u2y2; it is 85% of output and as-output items.

DMUjEoInput ItemsIntermediate ItemsOutput Items

Table 5.

Virtual weight percentage.


4. Discussion and conclusions

The objective of efficiency assessment is to identify weaknesses such that the appropriate steps to improve the entire system’s performance. This chapter introduces a two-phase procedure to evaluate the two-stage and network models with “free” links. This new model adopts SBM and considers not only the input and output slacks in the objective function but also the slacks of links. The resultant DEA scores provide completely information on how to project inefficient DMUs onto the DEA frontier for specific two-stage processes. Instead of the two conflicting roles that each link plays in existing models, each link plays a single role in the proposed two-phase process system in that it is either desirable or undesirable. The SBM model in this chapter counts the slacks associated with links in the efficiency scores, overcoming the hurdle. The bank case study takes the example on adjustment in the slacks and defines the best practice performance that the DMU under evaluation will need to attain to achieve the best efficiency. To achieve the best-practice efficiency, each DMU determines a set of weights for input, output, and link, where the links are designated as either “as-input” or “as-output”. Input and as-input measures reduce slacks, while output and as-output measures increase slacks to reach their targets on the production frontier. This study only introduces a two-stage procedure to assess the entire system. It also can be extended to more complex network processes, applied in series multistage, share resource (Chen et al. [21] and Liang et al. [22]), dynamic network DEA (Tone & Tsutsui [2] and Kao [13]), assurance region (Thompson et al. [23]), cone ratio model (Charnes et al. [24]), and virtual weight analysis models (Sarrico & Dyson [25]) in future research.



This research is supported by National Science Council of Taiwan, Republic of China, under the project NSC100-2221-E-009-065-MY3.


Conflict of interests

The authors declare that there is no conflict of interest regarding the publication of this paper.

© 2018 The Author(s). Licensee IntechOpen. This chapter is distributed under the terms of the Creative Commons Attribution 3.0 License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

How to cite and reference

Link to this chapter Copy to clipboard

Cite this chapter Copy to clipboard

Yu-cheng Liu (November 5th 2018). Two-Phase Network Data Envelopment Analysis: An Example of Bank Performance Assessment, Multi-Criteria Methods and Techniques Applied to Supply Chain Management, Valerio A. P. Salomon, IntechOpen, DOI: 10.5772/intechopen.74933. Available from:

chapter statistics

713total chapter downloads

More statistics for editors and authors

Login to your personal dashboard for more detailed statistics on your publications.

Access personal reporting

Related Content

This Book

Multi-Criteria Methods and Techniques Applied to Supply Chain Management

Edited by Valerio Salomon

Next chapter

Malmquist Index with Time Series to Data Envelopment Analysis

By Jhon Jairo Vargas Sánchez

Related Book

First chapter

Lean Supply Chain Practices and Performance in the Context of Malaysia

By Azman Daud and Suhaiza Zailani

We are IntechOpen, the world's leading publisher of Open Access books. Built by scientists, for scientists. Our readership spans scientists, professors, researchers, librarians, and students, as well as business professionals. We share our knowledge and peer-reveiwed research papers with libraries, scientific and engineering societies, and also work with corporate R&D departments and government entities.

More About Us