Socio-Economic Planning Sciences 45 (2011) 10e15

Contents lists available at ScienceDirect

Socio-Economic Planning Sciences

journal homepage: www.elsevier.com/locate/seps

The biennial Malmquist productivity change index

Jesús T. Pastor a, Mette Asmild b, *, C.A. Knox Lovell c
a
Centro de Investigación Operativa, Universidad Miguel Hernández de Elche, 03202 Elche, Alicante, Spain
b
Warwick Business School, University of Warwick, Coventry CV4 7AL, UK
c
CEPA, School of Economics, University of Queensland, Brisbane 4072, Australia

a r t i c l e i n f o a b s t r a c t

Article history: In this paper we introduce a new Malmquist productivity index that has three attractive features: it
Available online 1 October 2010 avoids linear programming infeasibilities under variable returns to scale, it allows for technical regress,
and it does not need to be recomputed when a new time period is added to the data set. The proposed
JEL Classification: index is compared to both the adjacent Malmquist index and the global Malmquist index in an empirical
C43 example, which highlights the drawbacks of the existing indexes compared to the proposed biennial
D24
Malmquist index.
O47
Our results show that 13% of the observations in the data set may have to be ignored due to infeasi-
Keywords:
bilities when decomposing the adjacent Malmquist index. Using only this reduced data set does at times
Malmquist productivity indices
lead to quite different results than those generated by applying the proposed biennial Malmquist index to
Infeasibilities
Technical regress the entire data set. The empirical example also shows that productivity change estimated between two
Productivity change decompositions time periods using the global Malmquist index change substantially when a third time period is added to
the data set, whereas the proposed biennial Malmquist index is immune to this problem.

2010 Elsevier Ltd. All rights reserved.

1. Introduction scores against the latter may result in linear programming infeasi-
bilities, and this has been a key argument posed against the use of
Malmquist productivity change indices, as defined in [1], have for this decomposition. The same infeasibilities problem is present,
the last three decades been the most commonly used indices to though not explicitly acknowledged, in [4,6], in which mixed-period
estimate productivity changes. Färe et al. [2] proposed to calculate efficiency scores are calculated relative to a variable returns to scale
Malmquist indices from DEA efficiency scores and first decomposed technology and used to calculate input mix or output mix effects. In
the constant returns to scale Malmquist index into its efficiency either case the data set itself will obviously reveal how much
change and technical change components; a two factor decompo- information is lost through infeasibilities (c.f. [7]).
sition which has been widely, though not universally, accepted. Recently two variable returns to scale Malmquist indices that
Their subsequent three-factor decomposition, in which the effi- avoid infeasibilities have been proposed. The first is the sequential
ciency change component is further decomposed into a scale change Malmquist index [8] that considers a sequence of technologies con-
and a pure technical efficiency change component has, however, structed from all data for all units and all periods up to and including
received a good deal of criticism for being internally inconsistent, the two periods being compared. The second is the global Malmquist
and was challenged in [3], though this decomposition in turn has index [9] that considers a single global technology constructed from
been criticised in [4] for not distinguishing between scale efficiency all data for all units and all periods of the sample. The main drawback
change and the input mix or output mix effects. For reviews of this of the sequential Malmquist index is that it precludes identification of
issue, including alternative decompositions, see [4,5]. technical regress, while the main drawback of the global Malmquist
In the decomposition in [3] the scale change effect is obtained by index is that adding additional time periods to a sample requires re-
considering both constant returns to scale and variable returns to calculation of all previous estimates, which therefore may change.
scale technologies. However estimating mixed-period efficiency These deficiencies have induced us to propose a new biennial
Malmquist index, which avoids infeasibilities, measures technical
regress as well as progress, and maintains previously computed
* Corresponding author.
productivity changes if a new time period is added to the data set.
E-mail addresses: jtpastor@umh.es (J.T. Pastor), mette.asmild@wbs.ac.uk The paper unfolds as follows: The biennial Malmquist index is
(M. Asmild), k.lovell@uq.edu.au (C.A.K. Lovell). introduced in Section 2 and decomposed in Section 3. Section 4

0038-0121/$ e see front matter
2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.seps.2010.09.001
 J.T. Pastor et al. / Socio-Economic Planning Sciences 45 (2011) 10e15 11

presents an empirical example that highlights the virtues of the Since we are using the biennial CRS technology, which includes
biennial index relative to its competitors, and Section 5 concludes. both the period t and period tþ1 technologies, we do not need to
resort to any geometric mean when defining (3).
2. The biennial Malmquist index The CRS benchmark technologies should be distinguished from
the best practice technologies allowing for variable returns to scale
In the following we consider output-oriented distance functions (VRS). “This convention enables it [the Malmquist index] to incor-
and Malmquist indices and a balanced panel of j ¼ 1,.,n producers porate the influence of scale economies as a departure of the best
in each of t ¼ 1,.,u time periods. Denote by ðx; yÞ˛<m s
þ  <þ the practice technology from the benchmark technology.” ([5], p. 440]).
input-output vector of a generic producer and by ðxj ; yj Þ˛<þ  <sþ
t t m To define VRS counterparts of the CRS constructs above, consider
the corresponding vector for a specific producer j in time period t. first the period t VRS technology defined as
For each t consider two benchmark technologies, the period t
Pn 8 
j ¼ 1 lj xj ; y 
t t
technology defined as Tct ¼ fðx; yÞ˛<mþs þ jx  
Pn <  X n Xn
j ¼ 1 lj yj ; lj  0; j ¼ 1; .; ng, and the technology associated with
t t t
mþs 
Tvt ¼ ðx; yÞ˛<þ x  ltj xtj ; y  ltj ytj ;
the subsequent period, Tctþ1 , defined similarly. Based on these two :  j¼1 j¼1
technologies the base t biennial technology TcBcan be defined as the 9
convex hull of the period t and period tþ1 technologies Xn =
TcB ¼ convfTCt ; TCtþ1 g. The subscript c in Tck ; k ¼ t; t þ 1 indicates
ltj ¼ 1; ltj  0; j ¼ 1; .; n
;
j¼1
that Tck exhibits constant returns to scale (CRS), i.e. lTck ¼ Tck for all
l > 0. Hence TcBalso satisfies CRS.
The only difference between Tct and Tvt is that the latter includes
The biennial technology corresponds to a DEA window analysis
the convexity constraint on the lambdas. Similarly the remaining
with a window width of two (see e.g. [10]). In the case of panel data
VRS technologies are easily defined and denoted by the subscript
consisting of two time periods, i.e. u ¼ 2, the biennial technology is
“v” rather than “c”. Hence the adjacent VRS Malmquist index is
identical to a pooled or meta-frontier technology also used to
given by
construct a global Malmquist index [9]. But in the general case of
more than two time periods (u > 2), a series of u1 overlapping   h  
biennial technologies exists for each pair-wise comparison of Mv xtj ; ytj ; xtþ1
j
; ytþ1
j
¼ Mvt xtj ; ytj ; xtþ1
j
; ytþ1
j
adjacent time periods. The biennial Malmquist index is defined  i1
 Mvtþ1 xtj ; ytj ; xtþ1 ; ytþ1
2
specifically for the adjacent time periods t and tþ1 since two j j
: (4)
adjacent time periods are sufficient to establish the desirable
properties of avoiding infeasibility, allowing technical regress, and and the biennial VRS Malmquist index is defined by
maintaining previous productivity calculations. Sufficiency of two  
time periods for these desirable properties does not preclude the   DBv xtþ1
j
; ytþ1
j
MvB xtj ; ytj ; xtþ1
j
; ytþ1
j
¼   : (5)
construction of a triennial Malmquist index. However the biennial DBv xtj ; ytj
Malmquist index is not transitive because it is constructed from
a series of overlapping two period technologies, and these tech- Finally note that a sufficient condition for MB to collapse to M is
nologies can differ. This drawback, however, is not uncommon; it is that Tt ¼ Ttþ1because, in that case, TB ¼ Tt. Moreover, if TB ¼ Tt then
shared by all Malmquist indices except the global index [9], which M ¼ Mt ¼ Mtþ1 ¼ MB for both constant and variable returns to scale
is transitive because it contains a single technology. technologies. A weaker sufficient condition is a Hicks neutrality
Based on the classic CRS output distance function for (x,y) property on Tt and Ttþ1. Geometrically, what in [11] is called
defined on the period t technology, Dtc ðx; yÞ ¼ minff > 0jðx; fy Þ˛Tct g, “implicit Hicks neutral technical change” means that Ttþ1 can be
the standard output-oriented adjacent period t Malmquist index for obtained from Tt through a parallel shift. In this case, TB equals Tt
producer j is given by and Ttþ1 and M ¼ Mt ¼ Mtþ1 (see [11,12]). An even weaker time
  neutrality condition which guarantees that M ¼ Mt ¼ Mtþ1 appears
  Dtc xtþ1
j
; ytþ1
j in [13].
Mct xtj ; ytj ; xtþ1
j
; ytþ1
j
¼   : (1) Fig. 1 provides an illustration of the biennial Malmquist index for
Dtc xtj ; ytj
a simple 1 input, 1 output example with three producers (A, B, C)
and the adjacent period tþ1 Malmquist index is defined similarly, observed in two time periods (t ¼ 1,2). The VRS and CRS frontiers for
using the output distance function defined on the technology for period 1 are indicated by dashed lines and the corresponding period
period tþ1, Dtcþ1. Mtcþ1 likely differs from Mtc, which leads to the 2 frontiers by the solid black lines. The biennial VRS frontier is
definition of the adjacent Malmquist productivity index [1], Mc, as
the geometric mean of Mtc and Mtcþ1:
  h  
Mc xtj ; ytj ; xtþ1
j ; ytþ1
j ¼ Mct xtj ; ytj ; xtþ1
j ; ytþ1
j
 i1
 Mctþ1 xtj ; ytj ; xtþ1 ; ytþ1
2
j j
: (2)

Similarly to the definition of Dtc, we define the biennial output

distance function, DBc , based on TcB instead of Tct, i.e.
DBc ðx; yÞ ¼ minff > 0jðx; fy Þ˛TcB g. We further define the biennial
CRS Malmquist index for producer j as
 
  DBc xtþ1
j
; ytþ1
j
McB xtj ; ytj ; xtþ1
j ; ytþ1
j ¼   : (3)
DBc xtj ; ytj
Fig. 1. Simple 1 input-1 output illustration.
12 J.T. Pastor et al. / Socio-Economic Planning Sciences 45 (2011) 10e15

 
indicated by the thick grey line and the biennial CRS frontier in this
Dtþ1
v xtþ1
j
; y tþ1
j
case coincides with that of period 2. Consider the period 2 producer McB ¼ MvB  SC B ¼ ECvB  TCvB  SC B ¼  
indicated by A2. In an output-oriented analysis the (mixed-period) Dtv xtj ; ytj
efficiency score for this observation relative to the VRS frontier for  .  
period 1 is undefined, and illustrates the infeasibilities occurring DBv xtþ1
j
; y tþ1
j
D tþ1 xtþ1 ; ytþ1
v j j
when using the decompositions [3] and [4]. But since the biennial   .  
DBv xtj ; ytj Dtv xtj ; ytj
VRS frontier envelops all observations from both time periods,
 .  
infeasibilities cannot occur and the biennial VRS output distance DBc xtþ1 ; y tþ1
D B xtþ1 ; ytþ1
j j v j j
function for A2 is well defined.   .   : ð9Þ
The biennial Malmquist index for A can now be calculated as the DBc xtj ; ytj DBv xtj ; ytj
ratio of the distances of A2 and A1 to the biennial frontier, which in the
output-oriented VRS case (and using slightly informal notation) is
given by MvB ðA1 ; A2 Þ ¼ DBv ðA2 Þ=DBv ðA1 Þ ¼ 1=0:857 ¼ 1:167; a value The corresponding decomposition of the adjacent Malmquist
which indicates the extent of (output-oriented) productivity growth index, as shown in [3], is
for producer A between periods 1 and 2, relative to the biennial VRS
frontier. In the CRS case the biennial Malmquist index for A is
Mc ¼ Mv  SC ¼ ECv  TCv  SC (10)
MCB ðA1 ; A2 Þ ¼ DBC ðA2 Þ=DBC ðA1 Þ ¼ 1=0:667 ¼ 1:5 where ECv ¼ ECBv ,

3. Decompositions 2    31
Dtv xtþ1
j
; y tþ1
j
D t xt ; yt
v j j
2

TCv ¼ 4    5 : (11)
3.1. Three-factor decomposition Dtþ1
v xtþ1
j
; ytþ1
j
Dtþ1
v xtj ; ytj

In what follows we derive the three-factor decomposition of [3] and

for the biennial Malmquist index. The third factor is the oft-con- 2  .  
tested scale change (SCB) component, and since McB ¼ MvB  SC B , Dtc xtþ1
j
; y tþ1
j
D t xtþ1 ; ytþ1
v j j
we obtain directly that SC ¼ 4  .  
Dtc xtj ; ytj Dtv xtj ; ytj
 .    .  31
DBc xtþ1 ; y tþ1
D B xtþ1 ; ytþ1 Dtþ1 xtþ1 ; ytþ1 Dtþ1
v xtþ1 ; ytþ1 2
McB j j v j j

c j

j
. 
j

j
5 :
SC B ¼ B ¼  .   : (6) (12)
Mv DBc xtj ; ytj DBv xtj ; ytj Dtþ1
c xtj ; ytj Dtþ1
v xtj ; ytj

The decomposition of the global Malmquist index is that of (9),

The scale change factor SCB is the ratio of the scale effects of but with the superscript B replaced by G, where G represents the
periods tþ1 and t. global technology TG, defined in [9] as TcG ¼ convfTc1 ; Tc2 ; .; Tcu g i.e.,
What remains is to decompose MBv into two factors, efficiency the convex hull of the technologies of all time periods.
change,ECBv , and technical change, TCBv . The efficiency change is, as
usual, defined as
3.2. Four-factor decomposition
 
Dtþ1
v xtþ1
j
; ytþ1
j Consider finally the four-factor decomposition in [4]. We replace
ECvB ¼   ; (7) the adjacent Malmquist index with the biennial Malmquist index to
Dtv xtj ; ytj avoid infeasibilities, generating
while the technical change factor corresponds to what is left over,
McB ¼ MvB  SC B ¼ ECvB  TCvB  SEC B  OMEB (13)
and so
where the change in scale efficiency is given by
 .  
2 .  
M B DBv xtþ1 ;ytþ1 DBv xtj ;ytj
TCvB ¼ vB ¼ 
j j
.   DBc xtþ1
j
; y t
j D B xtþ1 ; yt
v j j
ECv Dtþ1 xtþ1 ;ytþ1 Dtv xtj ;ytj SEC B ¼ 4  .  
v j j DBc xtj ; ytj DBv xtj ; ytj
 .    
 .  3 1
DBv xtþ1 ;ytþ1 Dtþ1 xtþ1 ;ytþ1 BPGB;tþ1
v xtþ1 ;ytþ1
¼
j

j
. 
v j

j
¼ 
j j
 : DBc xtþ1
j
; ytþ1
j
DBv xtþ1 j
; ytþ1
j
2

  .   5 : (14)
DBv xtj ;ytj Dtv xtj ;ytj BPGB;t
v xtj ;ytj
DBc xtj ; ytþ1
j
DBv xtj ; ytþ1 j
(8)
  and the output mix effect by
BPGB;tþ1
v xtþ1
j
measures the “best practice gap” between TvB
;ytþ1
j
  2  .  
and Tvtþ1 along the ray defined by xtþ1
j
;ytþ1
j
. Hence TCBv >1 [<1] DBc xtþ1
j
; ytþ1
j
DBv xtþ1j
; ytþ1
j
OMEB ¼ 4  .  
indicates that the best practice technology in period tþ1 along the DBc xtþ1 ; ytj DBv xtþ1 ; ytj
  j j
ray defined by xtþ1
j
;ytþ1
j
is closer to [further away from] the  .  31
biennial best practice technology than is the best practice tech- DBc xtj ; ytþ1
j
DBv xtj ; ytþ1
j
2

  .   5: (15)
nology in period t along the ray defined by ðxtj ;ytj Þ, i.e., it measures DBc xtj ; ytj DBv xtj ; ytj
technological progress [regress].
The complete three-factor decomposition of the biennial and thus the full four-factor decomposition of the biennial Malm-
Malmquist index is thus given by quist index is
 J.T. Pastor et al. / Socio-Economic Planning Sciences 45 (2011) 10e15 13

  0:5=0:571
McB ¼ ECvB TCvB SEC B OMEB McB C 1 ; C 2 ¼ 0:571=0:692 
0:667=0:692
 1  1
   .   0:5=0:75 0:333=0:5 2 0:333=0:5Þ 0:267=0:444Þ 2
Dtþ1 xtþ1 ;ytþ1 DBv xtþ1 ;ytþ1 Dtþ1 xtþ1 ;ytþ1    
v j j j j v j j 0:4=0:667 0:267=0:444 0:5=0:75 0:4=0:667
¼     .  
Dtv xtj ;ytj DBv xtj ;ytj Dtv xtj ;ytj
1 1
¼ 0:825  0:909  ½1:112  1:1082 ½1  12
2  .    .  3 1 ¼ 0:825  0:909  1:110  1 ¼ 0:833;
DBc xtþ1
j
;y t
j D B xtþ1 ;yt
v j j DB xtþ1 ;ytþ1
c j j
D B xtþ1 ;ytþ1
v j j
2

4  .     .   5
DBc xtj ;ytj DBv xtj ;ytj DBc xtj ;ytþ1 DBv xtj ;ytþ1 where the scale effect consists of scale efficiency change alone, as
j j
there is no possible change in output mix.
2  .    .  3 1 Alternatively if, for example, the observations in the example all
DBc xtþ1 ;ytþ1 DBv xtþ1 ;ytþ1 DBc xtj ;ytþ1 DBv xtj ;ytþ1 2
j j j j j j had values of 1 for a second output, except C1 with a value of 0.8 and
4  .     .   5
DBc xtþ1 ;ytj DBv xtþ1 ;ytj DBc xtj ;ytj DBv xtj ;ytj C2 with a value of 0.667, then the scale change component in the
j j
three-factor decomposition for observation C would be SC ¼ 1.00,
 
decomposing into SEC ¼ 0.8 and OME ¼ 1.25 in the four-factor
DBc xtþ1
j
;ytþ1
j ð16Þ
¼   : decomposition.
DBc xtj ;ytj
4. An empirical illustration

To illustrate the biennial Malmquist index and compare it with

3.3. Decompositions for the example in Fig. 1
the adjacent and global indices we reconsider the data set in [9],
which is a balanced panel of 93 US electricity generating firms in
In the simple example of Fig. 1, the three-factor decomposition
each of four years (1977, 1982, 1987, 1992). The firms use three
of the biennial CRS Malmquist index for A becomes
inputs, labor (FTE employees), fuel (BTUs of energy) and capital (a
  1 1=1 1=1 multilateral Törnqvist index) to generate a single output, electricity
McB A1 ; BA2 ¼ ECvB  TCvB  SC B ¼   (net generation in MWh).
1 0:857=1 0:667=0:857
¼ 1  1:167  1:286 ¼ 1:5; Table 1 summarizes the results of the biennial Malmquist index
estimated for each of the three overlapping time spans and decom-
and the four-factor decomposition is posed according to (16), though the three-factor decomposition (9)
  follows automatically since SCB ¼ SECB  OMEB ¼ SECB. Table 1 also
McB A1 ; A2 ¼ ECvB  TCvB  SEC B  OMEB shows the corresponding adjacent and global Malmquist indices in
 1 the final two columns, although these have not been decomposed
1 1=1 1=1 1=1 2
because numerous infeasibilities generate average values based on
¼   
1 0:857=1 0:667=0:857 0:667=0:857 fewer observations than for the biennial Malmquist index.
 1 The biennial index MBc and its two decompositions show that
1=1 0:667=0:857 2
  mean productivity change is improving through time, progressing
1=1 ð0:667=0:857Þ
from productivity decline to productivity growth. The technical
¼ 1  1:167  1:286  1 ¼ 1:5: change component exhibits the same pattern. The scale change
In this case there is no output mix effect since A has only one component hovers near unity in each period. Apart from the third
output, but there is a scale efficiency improvement from the period, the efficiency change component is the primary contributor
reduction in input use while maintaining the same output to productivity change.
production in both time periods. The adjacent Malmquist index Mc exhibits a similar, but more
For producer B the decompositions are extreme, pattern. The implausibly low average value of 0.43 from
  1977 to 1982 trails those of the global and the biennial Malmquist
1 0:8=1 0:444=0:8
McB B1 ; B2 ¼ ECvB  TCvB  SC B ¼   indices of 0.68 and 0.71 respectively. Moreover, as reported in [9],
1 1=1 0:556=1 the adjacent Malmquist index is the geometric mean of very
¼ 1  0:8  1 ¼ 0:8; different (geometric) mean values of MC1977 ¼ 0.71 and
M1982
C ¼ 0.26. Estimating productivity change from a common two
and
  year frontier rather than taking the geometric mean of two
McB B1 ;B2 ¼ ECvB  TCvB  SEC B  OMEB
 1  1 Table 1
1 0:8=1 0:833=1 0:667=0:8 2 0:667=0:8 0:667=0:8 2 Summary statistics for all 93 observations for the biennial Malmquist index and its
¼     
1 1=1 0:833=1 0:667=0:8 0:833=1 0:833=1 3- and 4-factor decompositions as well as the adjacent and global indices.

¼ 1  0:8  1  1 ¼ 0:8; MBc ECBv TCBv SCB SECB OMEB Mc MG

c

1977e1982, n ¼ 93 Min 0.28 0.50 0.29 0.60 0.60 1.00 0.22 0.28
as there is no scale effect since B uses the same input level in both Max 1.17 1.63 1.00 1.48 1.48 1.00 0.78 1.09
time periods. Mean 0.71 1.03 0.68 1.03 1.03 1.00 0.43 0.68
StDev 0.19 0.22 0.17 0.14 0.14 0.00 0.11 0.17
Finally, for producer C the decomposition is
1982e1987, n ¼ 93 Min 0.47 0.65 0.41 0.51 0.51 1.00 0.45 0.56
Max 1.68 3.01 1.11 1.68 1.68 1.00 1.77 1.86
  0:571 0:5=0:571 0:333=0:5
Mean 0.89 1.24 0.75 0.96 0.96 1.00 0.89 1.06
StDev 0.25 0.38 0.17 0.13 0.13 0.00 0.24 0.26
McB C 1 ; C 2 ¼  
0:692 0:667=0:692 0:4=0:667 1987e1992, n ¼ 93 Min 0.53 0.59 0.92 0.81 0.81 1.00 0.58 0.45
Max 2.00 1.67 1.52 1.30 1.30 1.00 2.15 2.32
¼ 0:825  0:909  1:111 ¼ 0:833 Mean 1.03 0.93 1.09 1.01 1.01 1.00 1.04 1.04
StDev 0.22 0.17 0.09 0.10 0.10 0.00 0.23 0.29
and
14 J.T. Pastor et al. / Socio-Economic Planning Sciences 45 (2011) 10e15

Table 2 properties may be more important than others, which in turn

Summary statistics for the adjacent and biennial Malmquist indices and their 3- should inform the choice of a productivity change index.
factor decompositions for the 81 observations for which the adjacent decomposition
is feasible in all time periods.
The main advantage of the biennial Malmquist index is that it
avoids the linear programming infeasibilities that occur when
MBc ECBv TCBv SCB Mc ECv TCv SC
attempting to decompose the adjacent Malmquist Index [2e4,6].
1977e1982, n ¼ 81 Min 0.28 0.50 0.29 0.64 0.22 0.50 0.13 0.70 We believe that the ability to tell a story about the contributions of
Max 1.16 1.63 1.00 1.27 0.78 1.63 0.70 1.94
various drivers to productivity change is an important part of the
Mean 0.72 1.05 0.67 1.02 0.44 1.05 0.38 1.10
StDev 0.18 0.22 0.16 0.12 0.10 0.22 0.12 0.24 exercise. We also believe that being forced to tell that story on the
1982e1987, n ¼ 81 Min 0.47 0.65 0.41 0.51 0.45 0.65 0.42 0.60 basis of a feasible subset of the data, not because of inadequacies in
Max 1.67 3.01 1.11 1.10 1.59 3.01 1.10 1.15 the data but due to a shortcoming of the analytical technique,
Mean 0.88 1.27 0.74 0.94 0.89 1.27 0.74 0.94 diminishes the credibility of the story. Consequently we believe
StDev 0.24 0.39 0.17 0.12 0.23 0.39 0.14 0.10
1987e1992, n ¼ 81 Min 0.62 0.59 0.99 0.81 0.58 0.59 0.97 0.82
that immunity to infeasible observations constitutes an important
Max 2.00 1.67 1.37 1.30 2.15 1.67 1.49 1.29 advantage of the proposed biennial index. The empirical illustra-
Mean 1.03 0.93 1.10 1.01 1.03 0.93 1.10 1.01 tion in Section 4 provides support for each of these beliefs.
StDev 0.22 0.18 0.08 0.10 0.23 0.18 0.08 0.09 Of the two alternatives that do not suffer from the problem of
infeasibilities, we have already argued that the sequential Malm-
quist index cannot capture technical regress, so in cases in which
consecutive estimates appears to give a more balanced, less vola- regress might occur, as for example in agricultural productivity
tile, and far more plausible, picture of productivity change. studies in which the impact of inclement weather shows up as
Using Eqs. (10)e(12) to decompose the adjacent index generates technical regress, a sequential index is not appropriate. We have
12 infeasibilities among the 93 utilities over the 1977e1982 period, also noted that when a new time period is added to a data, set the
leaving just 81 observations on which to base a decomposition. The global Malmquist index must be recomputed, a consequence of
results from using this reduced data set are shown in Table 2, in which is that estimated productivity change between two periods
which the biennial index is re-estimated and decomposed using can be influenced by observations from another period, a sort of
the 81 observations that are feasible for the adjacent index, which “relevance of irrelevant alternatives” property. We believe that
is also decomposed. A comparison with Table 1 reveals how the immunity to both of these problems constitutes another significant
story told by the biennial index can change when 13% of the advantage of the biennial index. The empirical illustration in
observations are sacrificed to infeasibilities in the adjacent index. Section 4 is silent on the first advantage, but it provides support for
For instance the maximum scale changes in the first two time the second.
periods are much higher in the full dataset (1.48 and 1.68) than The price paid for these advantages is the lack of transitivity;
what is found when only considering the reduced number of a drawback it shares with both the sequential index and the widely-
observations (1.27 and 1.10). This example highlights the potential used adjacent Malmquist index. The trade-off is thus between the
dangers of being forced to use a subset of observations, which is the theoretical property of transitivity, which favors the global index,
likely consequence of decomposing the adjacent index. and the empirical advantages of linear programming (which is used
Table 2 further reveals how large differences between the in the vast majority of Malmquist studies) feasibility, allowance for
biennial and adjacent Malmquist indices translate into differences technical regress, and insensitivity of productivity calculations to
between corresponding components. Over the 1977e1982 period the addition of time periods, a set of empirical advantages enjoyed
the mean technical change components are significantly different, only by the biennial index.
with a value of 0.67 for the biennial index but only 0.38 for the Researchers have been using linear programming techniques
adjacent index. And as we argue above, in this illustration the (and, on rare occasions, econometric techniques) to estimate and
biennial Malmquist index is the more credible, since the adjacent decompose the adjacent Malmquist productivity index for the
index is the geometric mean of two very different indexes. The better part of two decades. The breadth of empirical applications is
differences between the indices are less extreme in the other two enormous, ranging from time-space prism vertices to Asian brown
time periods because infeasibilities do not occur. The differences clouds to digestive disorders to the tapir in Honduras. A Google
that do exist are numerically small and attributable to minor search on “Malmquist productivity index” on September 20, 2010
differences in definitions of components. generated “about 18,900 results.” The adjacent index remains the
Finally, the global Malmquist index gives mean values for the index of choice, despite its lack of transitivity, its potential for
productivity changes of 0.71, 1.09 and 1.07. However if the produc- infeasibility, and its interpretation difficulties we mentioned in
tivity changes for the first two time periods are recalculated based Section 1. The sequential and global indices have not yet gained
on a global technology comprising the years 1977, 1982 and 1987 traction, due no doubt to their relative youth, and perhaps also to
alone, the mean values become 0.74 and 1.13 respectively, and thus their own drawbacks mentioned above.
adding the additional year of 1992 would mean having to recalcu- The proposed biennial Malmquist productivity index has no
late, and here significantly change, previously estimated results. known empirical weaknesses, and it can be extended to a trien-
nial index and further if desired. Its lack of transitivity is common
5. Discussion and conclusions to two of its three competitors, and the adjacent index has
flourished despite this drawback. We believe that the empirical
Taking a step back, one might question whether there is a need advantages of the biennial index, which are shared by no other
for yet another variant of the Malmquist index and whether the Malmquist index, warrant its careful consideration by empirical
biennial Malmquist index proposed here is, in fact, better than researchers.
those alternatives already existing in the literature. While we
believe strongly that the answer to the first of these questions is Acknowledgements
“yes”, the answer to the second is not as straightforward, as it
involves trade-offs among alternative desirable properties of Jesús T. Pastor thanks the Spanish Ministry of Science and
a productivity index. Each of the proposed indices has different Education for supporting this research through grants MTM2004-
properties and, depending on the specific situation, certain 07473 and MTM2009-10479 as well as CIO for financial support.
 J.T. Pastor et al. / Socio-Economic Planning Sciences 45 (2011) 10e15 15

References (UK). Prof. Pastor’s research fields include location science, banking and efficiency
analysis. He has served on the editorial review or advisory board of more than 20
international journals. He has authored or co-authored 9 books in various fields of
[1] Caves DW, Christensen LR, Diewert WE. The economic theory of index
mathematics and has published over 70 research papers. His research has appeared in
numbers and the measurement of inputs, outputs and productivity. Econo-
such journals as Operations Research, OMEGA, Journal of the Operational Research Society,
metrica 1982;50(6):1393e414.
Operations Research Letters, European Journal of Operations Research, Location Science,
[2] Färe R, Grosskopf S, Norris M, Zhang Z. Productivity growth, technical prog-
Environment and Planning, Economics Letters, Top, International Journal of Information
ress, and efficiency change in industrialized countries. The American
Technology and Decision Making, Annals of Operations Research, Journal of Productivity
Economic Review 1994;84(1):66e83.
Analysis, Economic Theory and European Finance Review. Prof. Pastor is past Vice-presi-
[3] Ray SC, Desli E. Productivity growth, technical progress, and efficiency change
dent and past President of the Spanish Statistical and Operations Research Society (SEIO).
in industrialized countries: comment. The American Economic Review
He is also past Vice-rector of Research of the University Miguel Hernandez de Elche, and
1997;87(5):1033e9.
actually is the Director of the Operations Research Centre of the same University. Finally,
[4] Balk BM. Scale efficiency and productivity change. Journal of Productivity
Prof. Pastor has been a member of the regional government of the Comunidad Valenciana
Analysis 2001;15(3):159e83.
during three years, serving as General Director for Research and Technological Transfer.
[5] Lovell CAK. The decomposition of Malmquist productivity indices. Journal of
Productivity Analysis 2003;20(3):437e58.
[6] Balk BM. The many decompositions of productivity change. Downloadable
from, www.rsm.nl/bbalk. Mette Asmild is Associate Professor of Operational Research in Warwick Business
[7] Xue M, Harker PT. Note: ranking DMUs with infeasible super-efficiency in DEA School, University of Warwick, UK after previous employments at the University of
models. Management Science 2002;48(5):705e10. Nottingham, UK and the Centre for Management of Technology and Entrepreneurship
[8] Shestalova V. Sequential Malmquist indices of productivity growth: an at the University of Toronto, Canada. She holds a Ph.D. from KVL, now University of
application to OECD industrial activities. Journal of Productivity Analysis Copenhagen, Denmark, and a BSc and MSc from Copenhagen Business School. She is
2003;19(2e3):211e26. a member of the centre for Management Science and Production Economics (MSAP),
[9] Pastor JT, Lovell CAK. A global Malmquist productivity index. Economics FOI, University of Copenhagen, Denmark and is an associate editor of the Journal of
Letters 2005;88(2):266e71. Productivity Analysis and OMEGA. She has published widely in the field of efficiency and
[10] Charnes A, Clark CT, Cooper WW, Golany B. A developmental study of data productivity analysis in journals such as the European Journal of Operational Research,
envelopment analysis in measuring the efficiency of maintenance units in the OMEGA and the Journal of Productivity Analysis, as well as specialised journals in
US air forces. Annals of Operations Research 1985;2:95e112. various application areas.
[11] Chambers RG, Färe R. Hicks’ neutrality and trade biased growth: a taxonomy.
The Journal of Economic Theory 1994;64(2):554e67.
[12] Färe R, Grosskopf S. Intertemporal production frontiers. Boston MA: Kluwer C. A. Knox Lovell is Honorary Professor and member of the Centre for Efficiency and
Academic Publishers; 1996. Productivity Analysis in the School of Economics at the University of Queensland,
[13] Pastor JT, Lovell CAK. Circularity of the Malmquist productivity index. Australia, and Emeritus Professor in the Department of Economics at the University of
Economic Theory 2007;33(3):591e9. Georgia, USA. He has served as Editor-in-Chief of Journal of Productivity Analysis, and he
has served on editorial boards of several other journals in economics, management
science and operations research. He has co-authored four books and co-edited six
Jesús T. Pastor is Professor of Statistics and Operations Research at the Universidad books, all in the area of productivity and efficiency analysis. He has authored or co-
Miguel Hernandez of Elche, Spain. He earned an MBA and a Ph.D. in mathematical authored over 100 papers in journals such as Management Science, Journal of Economic
sciences from Valencia University, Spain. He has been visiting researcher at the Theory, European Journal of Operational Research, Economic Journal, Economic Theory and
Universities of Georgia (USA), Toronto (Canada), Queensland (Australia) and Warwick Annals of Operations Research.