European Journal of Operational Research 185 (2008) 418–429

www.elsevier.com/locate/ejor

O.R. Applications

Efficiency decomposition in two-stage

data envelopment analysis: An application to
non-life insurance companies in Taiwan
a,* b
Chiang Kao , Shiuh-Nan Hwang
a
Department of Industrial and Information Management, National Cheng Kung University, Tainan, Taiwan, ROC
b
Graduate School of Management, Ming Chuan University, Taipei, Taiwan, ROC

Received 7 January 2006; accepted 30 November 2006

Available online 12 January 2007

Abstract

The efficiency of decision processes which can be divided into two stages has been measured for the whole process as
well as for each stage independently by using the conventional data envelopment analysis (DEA) methodology in order to
identify the causes of inefficiency. This paper modifies the conventional DEA model by taking into account the series rela-
tionship of the two sub-processes within the whole process. Under this framework, the efficiency of the whole process can
be decomposed into the product of the efficiencies of the two sub-processes. In addition to this sound mathematical prop-
erty, the case of Taiwanese non-life insurance companies shows that some unusual results which have appeared in the inde-
pendent model do not exist in the relational model. In other words, the relational model developed in this paper is more
reliable in measuring the efficiencies and consequently is capable of identifying the causes of inefficiency more accurately.
Based on the structure of the model, the idea of efficiency decomposition can be extended to systems composed of multiple
stages connected in series.
Ó 2007 Elsevier B.V. All rights reserved.

Keywords: Data envelopment analysis; Efficiency; Decomposition; Tandem system

1. Introduction

Since Charnes et al. (1978), data envelopment analysis (DEA) has been widely applied to measure the relative
efficiency of a set of decision making units (DMUs) which apply the same inputs to produce the same outputs.
The results indicate how efficient each DMU has performed as compared to other DMUs in converting inputs
to outputs. An issue which is of greater concern to the inefficient DMUs is what factors that cause the ineffi-
ciency, although it is obvious that either reducing inputs or increasing outputs will improve their performance.
To answer this question, much effort has been devoted to breaking down the overall efficiency into compo-
nents so that the sources of inefficiency can be identified. One type of decomposition focuses on the structure of

*
Corresponding author. Tel.: +886 6 2753396.
E-mail addresses: [email protected] (C. Kao), [email protected] (S.-N. Hwang).

0377-2217/$ - see front matter Ó 2007 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2006.11.041
 C. Kao, S.-N. Hwang / European Journal of Operational Research 185 (2008) 418–429 419

the DEA model. Banker et al. (1984) break the overall efficiency of a DMU into the product of scale efficiency
and technical efficiency. Byrnes et al. (1984) further separate the congestion effect from the technical efficiency.
Kao (1995) decomposes the overall efficiency into a weighted arithmetic mean of the efficiencies of individual
outputs. A similar decomposition from the input side is also derived. Another type of decomposition empha-
sizes the stages of the production process. The complicated production process is divided into sub-processes, in
that some intermediate products are the outputs of a sub-process on the one hand and the inputs of another
sub-process on the other hand. The works of Färe and Grosskopf (1996, 2000) and Seiford and Zhu (1999)
are some examples of this approach. In the former type of decomposition, there exists some mathematical rela-
tionship between the overall efficiency and the component efficiencies, while for the latter type there is no spe-
cific relationship between those two parts. The reason is because the sub-processes in the latter type are
considered as independent processes in calculating their efficiencies. The model for calculating the efficiencies
of the sub-processes does not reflect any relationship between the components and the whole system.
The simplest case of a complicated production process is a tandem system, in which the whole production
process is composed of two sub-processes connected in series. Seiford and Zhu (1999) divide a commercial
bank’s production process into the stages of profitability and marketability. The inputs of the bank production
process are employees, assets, and shareholders’ equity, which are also the inputs of the first stage. The outputs
of the bank production process are market value, total return on investments, and earnings per share, which are
also the outputs of the second stage. In addition to the inputs and outputs of the system, there are two interme-
diate products, revenues and profits, which are the outputs of the first stage as well as the inputs of the second
stage. The efficiencies of the first stage, second stage, and the whole production process are calculated via three
independent DEA models for 55 US commercial banks. Decomposition of the production process helps identify
the source of inefficiency. Zhu (2000) follows the same idea to analyze the financial efficiency of the best 500 com-
panies as ranked by Fortune. This two-stage concept has also been applied to measure the performance of mental
health care programs (Schinnar et al., 1990), education sector (Lovell et al., 1994), American Major League
Baseball teams (Sexton and Lewis, 2003), information technology (Chen and Zhu, 2004; Chen et al., 2006), etc.
The objective of this paper is to investigate efficiency decomposition in a two-stage production process
where the outputs of the first stage are the inputs of the second stage. Different from previous studies, which
treat the whole production process and the two sub-processes as independent, this paper takes the series rela-
tionship of the two sub-processes into account in measuring the efficiencies. We will show that the overall effi-
ciency is the product of the efficiencies of the two sub-processes. This mathematically sound property reflects
the physical linkage of the two sub-processes with the whole process in a tighter way. The efficiencies calcu-
lated from this relational two-stage DEA approach are more meaningful than those calculated from the inde-
pendent two-stage DEA approach. The non-life insurance companies in Taiwan, whose production process
resembles the two-stage process, are utilized to illustrate the whole idea. The results are compared to those
calculated from the independent two-stage approach to make suitable discussions.
The structure of this paper is organized as follows. In the next section, we develop the relational two-stage
DEA model for measuring the efficiencies of the whole process as well as the two sub-processes. In Section 3, we
use the non-life insurance companies in Taiwan as an example to calculate the efficiencies by using the relational
model. Finally, the results are compared to those calculated from the independent model to draw conclusions.

2. Relational two-stage DEA model

Denote Xij, i = 1, . . . , m and Yrj, r = 1, . . . , s as the ith input and rth output, respectively, of DMU
j, j = 1, . . . , n. The conventional DEA model for measuring the efficiency of DMU k under the assumption
of constant returns-to-scale is the CCR model (Charnes et al., 1978):
,
X s X
m
Ek ¼ max ur Y rk vi X ik
r¼1 i¼1
,
X
s X
m ð1Þ
s:t: ur Y rj vi X ij 6 1; j ¼ 1; . . . ; n;
r¼1 i¼1

ur ; vi P e; r ¼ 1; . . . ; s; i ¼ 1; . . . ; m;
420 C. Kao, S.-N. Hwang / European Journal of Operational Research 185 (2008) 418–429

where e is a small non-Archimedean number (Charnes et al., 1979; Charnes and Cooper, 1984). Each DMU
applies m inputs to produce s outputs, and Ek is the relative efficiency of DMU k, where Ek = 1 indicates effi-
ciency and Ek < 1 for inefficiency.
Now, suppose a production process is composed of a series of two sub-processes as depicted in Fig. 1. The
whole process uses m inputs Xik, i = 1, . . . , m to produce s outputs Yrk, r = 1, . . . , s. Different from the conven-
tional one-stage production process, here the production process is composed of two sub-processes with q
intermediate products Zpk, p = 1, . . . , q. Moreover, the intermediate products Zpk are the outputs of stage 1
as well as the inputs of stage 2. The conventional two-stage DEA study (e.g., Seiford and Zhu, 1999) is to
use Model (1) to measure the overall efficiency and the following Models (2a) and (2b) to measure the efficien-
cies of stage 1, E1k , and stage 2, E2k , respectively:
,
X q X
m
E1k ¼ max wp Z pk vi X ik
p¼1 i¼1
,
X
q X
m
s:t: wp Z pj vi X ij 6 1; j ¼ 1; . . . ; n; ð2aÞ
p¼1 i¼1

wp ; vi P e; p ¼ 1; . . . ; q; i ¼ 1; . . . ; m;
,
X
s X
q
E2k ¼ max ur Y rk wp Z pk
r¼1 p¼1
,
X
s X
q
s:t: ur Y rj wp Z pj 6 1; j ¼ 1; . . . ; n; ð2bÞ
r¼1 p¼1

ur ; wp P e; r ¼ 1; . . . ; s; p ¼ 1; . . . ; q:
These two models are essentially the same as Model (1). The efficiencies of the whole process and the two sub-
processes are calculated independently.
To link the two sub-processes with the whole process, a model must describe this series relationship between
the whole process and the two sub-processes. Consider DMU k. Denote ur ; vi , and wp as the multipliers that
DMU k has selected to calculate its overall efficiency Ek and sub-process efficiencies E1k and E2k . Then we have,
,
Xs X
m

Ek ¼ ur Y rk vi X ik 6 1;
r¼1 i¼1
,
X
q X
m
E1k ¼ wp Z pk vi X ik 6 1; ð3Þ
p¼1 i¼1
,
X
s X
q
E2k ¼ ur Y rk wp Z pk 6 1;
r¼1 p¼1

the overall efficiency is the product of the efficiencies of the two sub-processes:
Ek ¼ E1k  E2k : ð4Þ

production process DMU k

X 1k Z 1k Y1k
X 2k Z 2k Y2k
sub-process 1 sub-process 2
. . .
. . .
. . .
X mk Z qk Ysk

Fig. 1. A tandem system with inputs X, outputs Y, and intermediate products Z.

 C. Kao, S.-N. Hwang / European Journal of Operational Research 185 (2008) 418–429 421

Based on this concept, the way to calculate the overall efficiency Ek, taking into account the series relationship
of the two sub-processes, is to incorporate the ratio constraints of the two sub-processes into Model (1):
,
X s Xm
Ek ¼ max ur Y rk vi X ik
r¼1 i¼1
,
X
s X
m
s:t: ur Y rj vi X ij 6 1; j ¼ 1; . . . ; n;
r¼1 i¼1
,
X
q X
m ð5Þ
wp Z pj vi X ij 6 1; j ¼ 1; . . . ; n;
p¼1 i¼1
,
X
s X
q
ur Y rj wp Z pj 6 1; j ¼ 1; . . . ; n;
r¼1 p¼1

ur ; vi ; wp P e; r ¼ 1; . . . ; s; i ¼ 1; . . . ; m; p ¼ 1; . . . ; q:
The constraint set is the totality of those of Models (1), (2a) and (2b). Note that here we require the multipliers
associated with Zpj to be the same no matter whether it plays the role of output or input. The reason will be
clear from the following example.
Referring to Table 1, suppose there are three DMUs labeled as A, B, and C which use 2, 4, and 5 units of
input X to produce 1.5, 4, and 4 units of intermediate product Z, which in turn produce 1.5, 5, and 6 units of
output Y, respectively. The input–output production diagrams of the two sub-processes are depicted in Fig. 2
in a counterclockwise orientation. The right part shows sub-process 1, where input X is used to produce inter-
mediate product Z and the left part shows sub-process 2, where intermediate product Z is used to produce
output Y. Clearly, B is efficient in sub-process 1 and C is efficient in sub-process 2 because they lie on the pro-
duction frontiers OD and OE, respectively. The efficiency scores of the three DMUs for the two sub-processes
as calculated from Model (1) are shown in Table 1 under the columns of E1k and E2k . To measure the overall
efficiency Ek, we first calculate the expected output Y from input X indirectly via intermediate product Z. For
example, DMU A uses 2 units of X to produce 1.5 units of Z with an efficiency score 0.75. Had A operated
efficiently, it should have produced 2 units of Z as indicated by A 0 on the frontier OD in Fig. 2. Subsequently,
had A produced 2 units of Z, it would have produced 3 units of Y as indicated by A* on frontier OE in Fig. 2.
Hence, the overall efficiency of A is EA ¼ Y A =Y A ¼ ð3=2Þ=3 ¼ 1=2, which is also the product of E1A and E2A .

Table 1
An example with the efficiencies measured from different models
DMU Data Relational model CCR
X Z Y E1k E2k Ek Ek
A 2 3/2 3/2 3/4 2/3 1/2 3/5
B 4 4 5 1 5/6 5/6 1
C 5 4 6 4/5 1 4/5 24/25

Z
Sub-process 2 Sub-process 1 D
E 6
C
C
C B B
B B C
3
A
A
A A

Y X
5 O 5

Fig. 2. Derivation of the expected output Y from input X via intermediate product Z.
422 C. Kao, S.-N. Hwang / European Journal of Operational Research 185 (2008) 418–429

Similarly, B and C find their expected intermediate product via points B 0 and C 0 on frontier OD; along with
the expected amount of intermediate product, the expected outputs B* and C* are found on frontier OE. Thus,
we obtain the overall efficiency EB ¼ Y B =Y B ¼ 5=6 and EC ¼ Y C =Y C ¼ 6=7:5. This result is depicted in Fig. 3,
where the horizontal axis denotes input X and the vertical axis denotes output Y. The production frontier con-
structed from the relational two-stage model is the straight line OG. If the overall efficiency is measured by the
conventional DEA model, or the independent two-stage model, the frontier will be the straight line OF, and
the expected outputs are A°, B°, and C°, respectively. The efficiencies calculated from the independent model
are also shown in Table 1 with the heading ‘‘CCR’’ for comparison.
In this example we require the input of sub-process 2 to be the expected output of the first sub-process in
calculating the expected output of the whole process. When this concept is generalized to theP case of multiple
q
intermediate products, it requires the aggregated value of the intermediate products, which is p¼1 wp Z pj , to be
the same.
Model (5) is a linear fractional program which can be transformed into the following linear program:
Xs
Ek ¼ max ur Y rk
r¼1
X
m
s:t: vi X ik ¼ 1;
i¼1
X
s X
m
ur Y rj  vi X ij 6 0; j ¼ 1; . . . ; n;
r¼1 i¼1 ð6Þ
X
q X
m
wp Z pj  vi X ij 6 0; j ¼ 1; . . . ; n;
p¼1 i¼1

X
s X
q
ur Y rj  wp Z pj 6 0; j ¼ 1; . . . ; n;
r¼1 p¼1

ur ; vi ; wp P e; r ¼ 1; . . . ; s; i ¼ 1; . . . ; m; p ¼ 1; . . . ; q:
After the optimalPmultipliersPur ; vi , and wp are solved,
P Ps  the efficiencies
Pq are obtained subsequently as Ek ¼
s 1 q m 2
u
r¼1 r

Y rk ; E k ¼ w
p¼1 p

Z pk = v
i¼1 i

X ik , and E k ¼ u
r¼1 r Y rk = w
p¼1 p Z pk . Obviously, we have Ek ¼ E1k  E2k .

8
C* F

7 o
C
B*
6 C
o
5 B
B

3 A*
o
A
2
A
1

X
O 1 2 3 4 5 6

Fig. 3. Production frontiers constructed from the relational model and the independent model.
 C. Kao, S.-N. Hwang / European Journal of Operational Research 185 (2008) 418–429 423

Quite probably, the optimal multipliers solved from (6) may not be unique; consequently, the decomposi-
tion of Ek ¼ E1k  E2k would not be unique, either. This makes the comparison of either E1k or E2k among all
DMUs lack a common basis. One solution to this problem is to find the set of multipliers which produces
the largest E1k while maintaining the overall efficiency score at Ek calculated from Model (6). This idea can
be formulated as
X q
E1k ¼ max wp Z pk
p¼1
X
m
s:t: vi X ik ¼ 1;
i¼1
Xs X
m
ur Y rk  Ek vi X ik ¼ 0;
r¼1 i¼1
Xs X
m
ð7Þ
ur Y rj  vi X ij 6 0; j ¼ 1; . . . ; n;
r¼1 i¼1
X
q X m
wp Z pj  vi X ij 6 0; j ¼ 1; . . . ; n;
p¼1 i¼1
Xs Xq
ur Y rj  wp Z pj 6 0; j ¼ 1; . . . ; n;
r¼1 p¼1
ur ; vi ; wp P e; r ¼ 1; . . . ; s; i ¼ 1; . . . ; m; p ¼ 1; . . . ; q:
After E1k is calculated from the above model, the efficiency of the second stage is obtained as: E2k ¼ Ek =E1k .
Alternatively, if the efficiency of the second stage is of more concernPto the decision maker, one can measure
s
E2k first by replacing the objective function of Model (7) with
P r¼1 ur Y rk and the first constraint with
q 1 1 2
w Z
p¼1 p pk ¼ 1. In this case, E k is calculated subsequently as: E k ¼ E k =E k .
Model (6) has a dual of the following form:
!
Xm X q Xs
v w u
Ek ¼ min h  e si þ sp þ sr
i¼1 p¼1 r¼1
X
n X
n
s:t: hX ik  aj X ij  bj X ij  svi ¼ 0; i ¼ 1; . . . ; m;
j¼1 j¼1
X
n X
n
ð8Þ
bj Z pj  cj Z pj  swp ¼ 0; p ¼ 1; . . . ; q;
j¼1 j¼1
X
n X
n
aj Y rj þ cj Y rj  sur ¼ Y rk ; r ¼ 1; . . . ; s;
j¼1 j¼1

aj ; bj ; cj ; svi ; swp ; sur P 0; j ¼ 1; . . . ; n; r ¼ 1; . . . ; s; i ¼ 1; . . . ; m; p ¼ 1; . . . ; q:

Substituting aj + cj by kj and bj  cj by dj, we obtain kj + d j = aj + bj. The above model can be re-written as
!
Xm Xq X s
v w u
Ek ¼ min h  e si þ sp þ sr
i¼1 p¼1 r¼1
X
n
s:t: ðkj þ dj ÞX ij þ svi ¼ hX ik ; i ¼ 1; . . . ; m;
j¼1
Xn
ð9Þ
dj Z pj  swp ¼ 0; p ¼ 1; . . . ; q;
j¼1
X
n
kj Y rj  sur ¼ Y rk ; r ¼ 1; . . . ; s;
j¼1

kj ; svi ; swp ; sur P 0; j ¼ 1; . . . ; n; r ¼ 1; . . . ; s; i ¼ 1; . . . ; m; p ¼ 1; . . . ; q:

424 C. Kao, S.-N. Hwang / European Journal of Operational Research 185 (2008) 418–429

When all dj, j = 1, . . . , n are equal to zero, this model boils down to the conventional CCR model. If, on the
other hand, some dj are not zero, then the production frontier constructed from the observations will deviate
from the observations, rather than passing through the observations as in the conventional case. This devia-
tion explains why it is possible that none of the DMUs would have an efficiency score of 1.

3. Non-life insurance companies in Taiwan

Similar to other service industries, the non-life insurance industry provides services to their clients to gen-
erate profit. There are several studies which use the DEA technique to measure the managerial performance of
this industry (Fecher et al., 1993; Noulas et al., 2001). Notably, the profit is not earned from insurance services
alone. Non-life insurance companies use the insurance premiums acquired through the systems of agencies,
brokers, solicitors, etc. as capital for investment. Hence, the whole production process of the non-life insur-
ance industry can be divided into two sub-processes: premium acquisition and profit generation. The first
sub-process is characterized by marketing of the insurance, where clients are attracted to pay direct written
premiums and reinsurance premiums are received from other insurance companies. The second sub-process
is characterized by investment, where premiums are invested in a portfolio to earn profit. According to the
Insurance Law of Taiwan, Republic of China, the premiums can only be invested in limited areas. Of these,
bank deposits account for the largest portion, followed by marketable securities, real estate, and mortgage
loans. The non-life insurance industry has a typical two-stage structure of production process; therefore,
the methodology developed in the preceding section can be applied to measure the efficiencies of the whole
process and the two sub-processes.
In Taiwan, there are 24 non-life insurance companies. Hwang and Kao (2006) apply the two-stage DEA
model of Seiford and Zhu (1999) to measure the efficiencies of the whole process and the two sub-processes
independently. This paper will calculate these efficiencies in a relational manner. The inputs of the system,
which are also the inputs of the first stage, used by Hwang and Kao (2006) are:

Operation expenses (X1): salaries of the employees and various types of costs incurred in daily operation.
Insurance expenses (X2): expenses paid to agencies, brokers, and solicitors; and other expenses associated
with marketing the service of insurance.

The outputs of the system, which are also the outputs of the second stage, are:

Underwriting profit (Y1): profit earned from the insurance business.

Investment profit (Y2): profit earned from the investment portfolio.

There are also two intermediate products in the system, which are the outputs of the marketing sub-process
as well as the inputs of the investment sub-process:

Direct written premiums (Z1): premiums received from insured clients.

Reinsurance premiums (Z2): premiums received from ceding companies.

The efficiency of the first stage measures the performance in marketing the service of insurance while the
efficiency of the second stage measures the performance in generating profit from the premiums. The product
of the efficiencies of the two sub-processes is the efficiency of the whole process. Table 2 shows the inputs,
intermediate products, and outputs of the 24 non-life insurance companies in Taiwan. This set of data is taken
from Hwang and Kao (2006), and are the averages of 2001 and 2002 (Insurance Institute, 2003).
By applying Model (6), the overall efficiencies Ek of the 24 non-life insurance companies are calculated.
Since profit earning is of more concern to the insurance companies, we choose to measure E2k first and calculate
E1k via E1k ¼ Ek =E2k afterwards. The results are shown in the left half of Table 3 under the heading ‘‘Relational
two-stage model’’.
Notably, none of the 24 insurance companies perform efficiently in both stages. This is seen in the non-effi-
cient overall scores of all the companies, where the highest score is 0.767 occurring at Fubon (No. 5), followed
 C. Kao, S.-N. Hwang / European Journal of Operational Research 185 (2008) 418–429 425

Table 2
Inputs (X), intermediate products (Z), and outputs (Y) of the 24 non-life insurance companies in Taiwan
Operation Insurance Direct written Reinsurance Underwriting Investment
expenses (X1) expenses (X2) premiums (Z1) premiums (Z2) profit (Y1) profit (Y2)
1. Taiwan Fire 1,178,744 673,512 7,451,757 856,735 984,143 681,687
2. Chung Kuo 1,381,822 1,352,755 10,020,274 1,812,894 1,228,502 834,754
3. Tai Ping 1,177,494 592,790 4,776,548 560,244 293,613 658,428
4. China 601,320 594,259 3,174,851 371,863 248,709 177,331
Mariners
5. Fubon 6,699,063 3,531,614 37,392,862 1,753,794 7,851,229 3,925,272
6. Zurich 2,627,707 668,363 9,747,908 952,326 1,713,598 415,058
7. Taian 1,942,833 1,443,100 10,685,457 643,412 2,239,593 439,039
8. Ming Tai 3,789,001 1,873,530 17,267,266 1,134,600 3,899,530 622,868
9. Central 1,567,746 950,432 11,473,162 546,337 1,043,778 264,098
10. The First 1,303,249 1,298,470 8,210,389 504,528 1,697,941 554,806
11. Kuo Hua 1,962,448 672,414 7,222,378 643,178 1,486,014 18,259
12. Union 2,592,790 650,952 9,434,406 1,118,489 1,574,191 909,295
13. Shingkong 2,609,941 1,368,802 13,921,464 811,343 3,609,236 223,047
14. South China 1,396,002 988,888 7,396,396 465,509 1,401,200 332,283
15. Cathay 2,184,944 651,063 10,422,297 749,893 3,355,197 555,482
Century
16. Allianz 1,211,716 415,071 5,606,013 402,881 854,054 197,947
President
17. Newa 1,453,797 1,085,019 7,695,461 342,489 3,144,484 371,984
18. AIU 757,515 547,997 3,631,484 995,620 692,731 163,927
19. North 159,422 182,338 1,141,950 483,291 519,121 46,857
America
20. Federal 145,442 53,518 316,829 131,920 355,624 26,537
21. Royal & 84,171 26,224 225,888 40,542 51,950 6491
Sunalliance
22. Asia 15,993 10,502 52,063 14,574 82,141 4181
23. AXA 54,693 28,408 245,910 49,864 0.1 18,980
24. Mitsui 163,297 235,094 476,419 644,816 142,370 16,976
Sumitomo
Mean 1,544,215 828,963 7,832,893 667,964 1,602,873 477,733

by 0.760, occurring at Union (No. 12). For stage 1, the sub-process of insurance marketing, there are three
companies, Union (No. 12), Cathay Century (No. 15), and North America (No. 19), which perform efficiently.
For stage 2, the sub-process of profit earning, there are two companies, namely, Tai Ping (No. 3) and Asia
(No. 22), which are efficient. Since the overall efficiency Ek is the product of the efficiencies of the first stage
E1k and the second stage E2k , every Ek is no greater than its corresponding E1k and E2k . The last row of Table 3
shows the averages of these three measures. Another point to be noted is that most companies have a smaller
E2k than E1k , even though we calculate E2k first and E1k afterwards in this example. Only Tai Ping (No. 3), Fubon
(No. 5), and Asia (No. 22) have a smaller E1k than E2k . A Wilcoxon’s matched-pairs signed-ranks test (Daniel,
1978) confirms that the efficiency of the first stage is higher than the second stage in a statistical sense. This
indicates that the low efficiency score of the whole production process is mainly due to the low efficiency score
of the second stage—the profit earning process.
Since Ek is always less than or equal to E1k and E2k , there is not much meaning in looking at the efficiency scores
themselves. Rather, it will be more informative to look at the ranks of the efficiency scores, which are the num-
bers in parentheses in Table 3. When ties occur, each tied value is assigned the mean of the rank positions for
which it is tied. Most companies have similar ranks in Ek ; E1k , and E2k , which implies that the performance of the
whole process is evenly attributed to the performance of the two sub-processes. However, there are still several
companies which have large difference among these ranks. A large difference reveals the source that causes the
inefficiency of the whole process. For example, Tai Ping (No. 3), Fubon (No. 5), Newa (No. 17), and Asis (No.
22) perform unsatisfactorily in stage 1 (as compared to stage 2) and Central (No. 9) and North America (No. 19)
426 C. Kao, S.-N. Hwang / European Journal of Operational Research 185 (2008) 418–429

Table 3
Efficiency measures and their ranks (in parentheses) of the relational model and independent model for the 24 non-life insurance
companies in Taiwan
Company Relational two-stage model Independent two-stage model
Ek E1k E2k Ek E1k E2k
1. Taiwan Fire 0.699 (3) 0.993 (6) 0.704 (5) 0.984 (6) 0.993 (7) 0.713 (7)
2. Chung Kuo 0.625 (5) 0.998 (5) 0.626 (6) 1.000 (2.5) 0.998 (6) 0.627 (10)
3. Tai Ping 0.690 (4) 0.690 (16) 1.000 (1.5) 0.988 (5) 0.690 (23) 1.000 (2.5)
4. China Mariners 0.304 (15) 0.724 (15) 0.420 (13) 0.488 (14) 0.724 (21) 0.432 (16)
5. Fubon 0.767 (1) 0.831 (12) 0.923 (3) 1.000 (2.5) 0.838 (13) 1.000 (2.5)
6. Zurich 0.390 (12) 0.961 (7) 0.406 (17) 0.594 (13) 0.964 (8) 0.406 (18)
7. Taian 0.277 (17) 0.671 (18) 0.412 (15) 0.470 (17) 0.752 (16) 0.538 (13)
8. Ming Tai 0.275 (18) 0.663 (20) 0.415 (14) 0.415 (19) 0.726 (19) 0.511 (15)
9. Central 0.223 (20) 1.0  e (4) 0.223 (24) 0.327 (22) 1.000 (3) 0.292 (23)
10. The First 0.466 (9) 0.862 (10) 0.541 (10) 0.781 (10) 0.862 (11) 0.674 (9)
11. Kuo Hua 0.164 (23) 0.647 (21) 0.253 (23) 0.283 (23) 0.741 (18) 0.327 (22)
12. Union 0.760 (2) 1.000 (2) 0.760 (4) 1.000 (2.5) 1.000 (3) 0.760 (6)
13. Shingkong 0.208 (21) 0.672 (17) 0.309 (21) 0.353 (20) 0.811 (14) 0.543 (12)
14. South China 0.289 (16) 0.670 (19) 0.431 (12) 0.470 (16) 0.725 (20) 0.518 (14)
15. Cathay Century 0.614 (6) 1.000 (2) 0.614 (7) 0.979 (7) 1.000 (4) 0.705 (8)
16. Allianz President 0.320 (14) 0.886 (9) 0.362 (18) 0.472 (16) 0.907 (10) 0.385 (19)
17. Newa 0.360 (13) 0.628 (22) 0.574 (9) 0.635 (11) 0.723 (22) 1.000 (2.5)
18. AIU 0.259 (19) 0.794 (13) 0.326 (19) 0.427 (18) 0.794 (15) 0.374 (20)
19. North America 0.411 (11) 1.000 (2) 0.411 (16) 0.822 (9) 1.000 (3) 0.416 (17)
20. Federal 0.547 (8) 0.933 (8) 0.586 (8) 0.935 (8) 0.933 (9) 0.901 (5)
21. Royal & Sunalliance 0.201 (22) 0.732 (14) 0.274 (22) 0.333 (21) 0.751 (17) 0.280 (24)
22. Asia 0.590 (7) 0.590 (23) 1.000 (1.5) 1.000 (2.5) 0.590 (24) 1.000 (2.5)
23. AXA 0.420 (10) 0.843 (11) 0.499 (11) 0.599 (12) 0.851 (12) 0.560 (11)
24. Mitsui Sumitomo 0.135 (24) 0.429 (24) 0.314 (20) 0.257 (24) 1.000 (3) 0.335 (21)
Mean 0.416 0.801 0.516 0.651 0.849 0.596

perform unsatisfactorily in stage 2 (as compared to stage 1). Decomposing the overall efficiency into the product
of two component efficiencies helps a company identify the sub-process that causes inefficiency.
Logically, the rank of Ek for a company should lie either between the ranks of E1k and E2k or in the neigh-
borhood of E1k or E2k because the performance of the whole process is the aggregated performance of the two
sub-processes. For the 24 companies, 11 lie between the ranks of E1k and E2k , and the remaining lie in the neigh-
borhood of 2 ranks of either E1k or E2k . This conforms to intuition.

4. Relational model vs independent model

For the same data set contained in Table 2, Hwang and Kao (2006) have calculated the overall efficiencies,
stage 1 efficiencies, and stage 2 efficiencies independently by using the conventional CCR model. The results
are reported in the right half of Table 3 under the heading ‘‘Independent two-stage model’’. There are four
companies, Chung Kuo (No. 2), Fubon (No. 5), Union (No. 12), and Asia (No. 22), which perform efficiently
in the whole production process. Interestingly, none of them performs efficiently in both sub-processes. This is
different from the rationale of the relational model. Of these four companies, Chung Kuo (No. 2) performs
efficiently in neither sub-process, Union (No. 12) performs efficiently only in the first sub-process, and Fubon
(No. 5) and Asia (No. 22) perform efficiently only in the second sub-process. On the other hand, there are also
companies which perform efficiently in either sub-process 1, i.e., Central (No. 9), Union (No. 12), Cathay Cen-
tury (No. 15), North America (No. 19), and Mitsui Sumitomo (No. 24), or sub-process 2, i.e., Tai Ping (No.
3), Fubon (No. 5), Newa (No. 17), and Asia (No. 22), yet their overall efficiency is not 1. Since in the inde-
pendent model Ek ; E1k and E2k are calculated independently, Ek is not necessarily smaller than either E1k or
E2k as in the relational model. As a matter of fact, the average of Ek is greater than that of E2k (referring to
the last row of Table 3). A Wilcoxon’s test shows that there is no significant difference both between Ek
 C. Kao, S.-N. Hwang / European Journal of Operational Research 185 (2008) 418–429 427

and E1k , and between Ek and E2k . Nevertheless, E1k is significantly greater than E2k . The last conclusion is the
same as that obtained from the relational model.
To investigate the difference between Ek calculated from the relational model and the independent model, it
is more appropriate to compare the ranks rather than the numerical scores because the relational model has
two more sets of constraints which result in smaller efficiency score for every company. The rankings of the
two models as shown in the second and fifth columns of Table 3 are quite similar. The largest difference is 4.5,
occurring at Asia (No. 22); the second largest difference is 3, occurring at Taiwan Fire (No. 1); and the third
largest difference is 2.5, occurring at Chung Kuo (No. 2). For all other 21 companies, their differences are less
than or equal to 2. A Spearman’s rank correlation test (Daniel, 1978) indicates that Ek from the two models
are highly correlated: a calculated coefficient of 0.972 as opposed to the tabulated value of 0.607 (n = 24,
a = 0.001). In other words, the independent model or the relational model produces Ek scores of almost
the same rankings.
The rankings of E2k calculated from the two models are also close. The largest difference occurs at Shing-
kong (No. 13), with a rank difference of 9. The second largest difference occurs at Newa (No. 17), with a rank
difference of 6.5. The third largest difference occurs at Chung Kuo (No. 2), with a rank difference of 4. There
are two companies, China Mariners (No. 4) and Federal (No. 20), which have a rank difference of 3. For the
remaining 19 companies, their rank differences are less than or equal to 2. The Spearman coefficient calculated
for this case is 0.918, which is greater than the tabulated value of 0.607, indicating a strong correlation.
The rankings of E1k calculated from the two models are not so close as the case of Ek or E2k . Mitsui Sumi-
tomo (No. 24) is the extreme case, while it is considered the worst company by the relational model, it is con-
sidered as efficient by the independent model. The second largest difference occurs at Tai Ping (No. 3), with a
rank difference of 7. The third largest difference occurs at China Mariners (No. 4), with a rank difference of 6.
There are three companies, Kuo Hua (No. 11), Shingkong (No. 13), and Royal & Sunalliance (No. 21), which
have a rank difference of 3. For the remaining 18 companies, their rank differences are less than or equal to 2.
Despite these six cases of large difference in ranks, the Spearman coefficient of 0.750 still indicates a strong
correlation between the rankings of E1k calculated from the two models.
By comparing the rankings of the efficiencies of the three processes obtained from the independent model,
one will find seven companies whose ranks are quite different. Of these, Central (No. 9), North America (No.
19), and Mitsui Sumitomo (No. 24) have much higher ranks for sub-process 1 than for sub-process 2; and Tai
Ping (No. 3), Fubon (No. 5), Newa (No. 17), and Asia (No. 22) have much lower ranks for sub-process 1 than
for sub-process 2. The dispersion of the ranks of Ek ; E1k , and E2k in this model is wider than that of the relational
model. Usually the rank of Ek for a company is between those of E1k and E2k , or in their neighborhood. In this
model there are 13 companies whose ranks of Ek lie between the ranks of E1k and E2k , and 8 lie in the neigh-
borhood of 2 ranks of either E1k or E2k . For the remaining 3 companies, Chung Kuo (No. 2) lies in the neigh-
borhood of 3.5 ranks. It performs efficiently in the whole process, yet its first sub-process is ranked sixth and
its second sub-process is ranked tenth. Shingkong (No. 13) lies in the neighborhood of 6 ranks. It has a four-
teenth-ranked first sub-process and twentieth-ranked second sub-process, yet its whole process is ranked twen-
tieth. Mitsui Sumitomo (No. 24) lies in the neighborhood of 3 ranks. It has a fifth-ranked first sub-process and
twenty-first-ranked second sub-process, yet its whole process is ranked the lowest. While the whole process is
composed of the two sub-processes, the results of these three companies are against expectation.
Of the three unusual cases, Mitsui Sumitomo (No. 24) is an interesting one, worth further discussing. Its
ranks in Ek ; E1k and E2k are 24, 24, and 20, respectively, for the relational model; and 24, 3, and 21, respectively,
for the independent model. While the ranks in Ek and E2k obtained from the two models are almost identical,
the ranks in E1k are reversed. The multipliers u, v, and w obtained from the two models may give some clues to
this phenomenon. Table 4 shows the optimal multipliers associated with this company in calculating different
efficiencies. For the independent model, when Y1 and Y2 are considered as the direct outputs of inputs X1 and
X2, the importance of X1 and X2 as indicated by the multipliers v1 and v2 in the first row of Table 4 are
6.124E6 and e, respectively, where the importance of X2 is negligible. The efficiency of the whole process,
0.2572, is the ratio of the aggregated output u 0 Y = 0.257 to the aggregated input v 0 X = 1. However, when
Y1 and Y2 are considered as the indirect outputs of inputs X1 and X2 via intermediate products Z1 and Z2,
the importance of X1 and X2 as indicated by the multipliers v1 and v2 in the second row of Table 4 are
e and 4.254E6, respectively, where, contradictorily, the importance of X2 becomes significant while the
428 C. Kao, S.-N. Hwang / European Journal of Operational Research 185 (2008) 418–429

Table 4
Optimal multipliers associated with company 24 in calculating efficiencies
Inputs (X) Intermediate products (Z) Outputs (Y)
v1 v2 v0X w1 w2 w0Z u1 u2 u0Y
Independent model
E24 6.124E6 e 1 0.735E6 0.981E6 0.257
E124 e 4.254E6 1 e 1.550E6 1
E224 2.099E6 e 1 0.568E6 14.974E6 0.335
Relational model
E24 6.124E6 e 1 0.834E6 0.048E6 0.429 0.233E6 5.991E6 0.135

importance of X1 becomes negligible. The efficiency of the first sub-process, 1, is the ratio of the aggregated
output w 0 Z = 1 to the aggregated input v 0 X = 1. It is also interesting to note that when Z1 and Z2 are
considered as the outputs of X1 and X2, the importance of Z1 is negligible, as indicated by the value e in
the second row of Table 4. However, when Z1 and Z2 are considered as the inputs of Y1 and Y2, the impor-
tance of Z2 becomes negligible, as indicated by its value e in the third row of Table 4. This seems to be another
contradictory point.
For the relational model, the sets of multipliers obtained from calculating E24 and E224 are the same; there-
fore, it is only reported once in the last row of Table 4. This set of multipliers shows that when the intermediate
products Z1 and Z2 are taken into account, the importance of X1 and X2 are 6.124E6 and e, respectively. The
importance of Z1 and Z2 are consistently 0.834E6 and 0.048E6, respectively, no matter whether they play
the role of input or output. The efficiency of the whole process is the ratio of u 0 Y = 0.135 to v 0 X = 1, for the
first sub-process it is w 0 Z/v 0 X = 0.429, and for the second sub-process it is u 0 Y/w 0 Z = 0.135/0.429 = 0.314.
This result gives a better explanation for a production system with intermediate products.

5. Conclusion

The objective of efficiency measurement is to detect the weak areas so that appropriate effort can be devoted
to improve performance. When a production system can be separated into two sub-processes, several studies
find that in addition to calculating the efficiency of the whole system by using the conventional DEA model,
the efficiencies of the two sub-processes can also be calculated to identify the source that causes the inefficiency
of the whole system. A deficiency of these studies is that the efficiencies of the whole process and the two sub-
processes are calculated independently, without taking into account the fact that the outputs of the first sub-
process are the inputs of the second sub-process.
In this paper the series relationship of the two sub-processes is accounted for in developing the new model.
To avoid the possibility of multiple solutions which would cause bias in comparison, the efficiency of the first
sub-process is maximized in a second stage under the constraint that the overall efficiency score is maintained
at the same level. Under this construct, the overall efficiency is the product of the efficiencies of the two sub-
processes. This mathematical relationship between the overall efficiency and the component efficiencies suit-
ably describes people’s expectation regarding the physical relationship of the whole process and the two
sub-processes. The non-life insurance case described in Hwang and Kao (2006) is used to illustrate whether
the independent two-stage model or the relational two-stage model better describes the performance of the
whole process and the sub-processes. The results show that the independent model may produce unusual
results for several companies while the relational model always produces meaningful results for all companies.
This leads to the conclusion that the relational model of this paper not only suitably describes the physical
relationship between the whole process and the component sub-processes, but also produces reliable results
in efficiency measurement.
Finally, there are systems which are composed of more than two sub-processes (e.g., Amado, 2003). The
formulation of the relational two-stage model shows that it can be extended to systems of multiple stages
connected in series, and the efficiency of the whole process is the product of the efficiencies of individual
sub-processes. However, a requirement for this relationship to hold is that all outputs of one sub-process must
 C. Kao, S.-N. Hwang / European Journal of Operational Research 185 (2008) 418–429 429

be the inputs of the next sub-process. If there is any output of a sub-process which is not the input of the next
sub-process or if there is any input of a sub-process which is not the output of its preceding sub-process, then
this relationship may not hold. How to relax this requirement to develop a more general model with sound
mathematical properties is a direction for future research.

Acknowledgement

This research is supported by the National Science Council, Republic of China, under contract number:
NSC95-2416-H-006-026-MY3. The authors are grateful to Dr. Wen-Shiang Wu for his help in calculating
the efficiencies.

References

Amado, C.F., 2003. Exploring the Use of Data Envelopment Analysis for Evaluation in Primary Health Care. Ph.D. thesis, The
University of Warwick.
Banker, R.D., Charnes, A., Cooper, W.W., 1984. Some models for estimating technical and scale efficiencies in data envelopment analysis.
Management Science 30, 1078–1092.
Byrnes, P., Färe, R., Grosskopf, S., 1984. Measuring productive efficiency: An application to Illinois strip mines. Management Science 30,
671–681.
Charnes, A., Cooper, W.W., 1984. The non-Archimedean CCR ratio for efficiency analysis: A rejoinder to Boyd and Färe. European
Journal of Operational Research 15, 333–334.
Charnes, A., Cooper, W.W., Rhodes, E., 1978. Measuring the efficiency of decision making units. European Journal of Operational
Research 2, 429–444.
Charnes, A., Cooper, W.W., Rhodes, E., 1979. Short communication: Measuring the efficiency of decision making units. European
Journal of Operational Research 3, 339.
Chen, Y., Zhu, J., 2004. Measuring information technology’s indirect impact on firm performance. Information Technology and
Management Journal 5, 9–22.
Chen, Y., Liang, L., Yang, F., Zhu, J., 2006. Evaluation of information technology investment: A data envelopment analysis approach.
Computers and Operations Research 33, 1368–1379.
Daniel, W.W., 1978. Applied Nonparametric Statistics. Houghton Mifflin, Boston.
Färe, R., Grosskopf, S., 1996. Productivity and intermediate products: A frontier approach. Economics Letters 50, 65–70.
Färe, R., Grosskopf, S., 2000. Network DEA. Socio-Economic Planning Sciences 34, 35–49.
Fecher, F., Kessler, D., Pestieau, P., 1993. Productive performance of the French insurance industry. Journal of Productivity Analysis 4,
77–93.
Hwang, S.N., Kao, T.L., 2006. Measuring managerial efficiency in non-life insurance companies: An application of two-stage data
envelopment analysis technique. International Journal of Management 23, 699–720.
Insurance Institute of the ROC, 2003. Non-Life Insurance Review, 2001–2002. Taipei, Taiwan.
Kao, C., 1995. Some properties of Pareto efficiency under the framework of data envelopment analysis. International Journal of Systems
Science 26, 1549–1558.
Lovell, C., Walters, L., Wood, L., 1994. Stratified models of education production using modified DEA and regression analysis. In:
Charnes, A., Cooper, W., Lewin, A., Seiford, L. (Eds.), Data Envelopment Analysis: Theory, Methodology and Applications. Kluwer
Academic Publishers, London, pp. 329–351.
Noulas, A.G., Hatzigayios, T., Lazaridis, J., Lyroudi, K., 2001. Non-parametric production frontier approach to the study of efficiency of
non-life insurance companies in Greece. Journal of Financial Management and Analysis 14, 19–26.
Schinnar, A.P., Kamis-Gould, E., Delucia, N., Rothbard, A.B., 1990. Organizational determinants of efficiency and effectiveness in mental
health partial care Programs. Health Services Research 25, 387–420.
Seiford, L.M., Zhu, J., 1999. Profitability and marketability of the top 55 US commercial banks. Management Science 45, 1270–1288.
Sexton, T.R., Lewis, H.F., 2003. Two-stage DEA: An application to major league baseball. Journal of Productivity Analysis 19, 227–249.
Zhu, J., 2000. Multi-factor performance measure model with an application to Fortune 500 companies. European Journal of Operational
Research 123, 105–124.