Professional Documents
Culture Documents
net/publication/223840142
CITATIONS READS
517 1,447
3 authors, including:
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Ca Knox Lovell on 09 September 2018.
a
Agricultural Economics Research Institute (LEI-DLO), P.O. Box 29703, 2502LS The Hague, The Netherlands
b
University of Georgia, Athens, GA 30602, USA
c
Wageningen Agricultural University, Wageningen, The Netherlands
Abstract
The objective of this paper is to estimate comprehensive environmental eciency measures for Dutch dairy farms.
The environmental eciency scores are based on the nitrogen surplus, phosphate surplus and the total (direct and
indirect) energy use of an unbalanced panel of dairy farms. We de®ne environmental eciency as the ratio of minimum
feasible to observed use of multiple environmentally detrimental inputs, conditional on observed levels of output and
the conventional inputs. We compare two methods for the calculation of eciency; namely Stochastic Frontier Analysis
(SFA) and Data Envelopment Analysis (DEA). This paper reveals the strengths and weaknesses for estimating envi-
ronmental eciency of the methods applied. Both SFA and DEA can estimate environmental eciency scores. The
mean technical eciency scores (output-oriented, SFA 89%, DEA 78%) and the mean comprehensive environmental
eciency scores (SFA 80%, DEA 52%) dier between the two methods. SFA allows hypothesis testing, and the
monotonicity hypothesis is rejected for the speci®cation including phosphate surplus. DEA can calculate environmental
eciency scores for all speci®cations, because regularity is imposed in this method. Ó 2000 Elsevier Science B.V. All
rights reserved.
Keywords: Environment; Agriculture; Eciency; Stochastic Frontier Analysis; Data Envelopment Analysis
1. Introduction
0377-2217/00/$ - see front matter Ó 2000 Elsevier Science B.V. All rights reserved.
PII: S 0 3 7 7 - 2 2 1 7 ( 9 9 ) 0 0 2 1 8 - 0
288 S. Reinhard et al. / European Journal of Operational Research 121 (2000) 287±303
agriculture has become the major objective of the to calculate shadow prices of the undesirable out-
Dutch agricultural policy. To achieve a competitive puts. Ball et al. (1994) and Tyteca (1997) provided
agriculture, farms have to apply marketable inputs empirical applications of the DEA model pro-
as eciently as possible, and to create environment- posed by Fare et al. (1989). Hetemaki (1996) also
friendly agriculture they have to deal eciently treated environmental eects as undesirable out-
with the environment. This raises the aggregate puts, and used econometric techniques to estimate
questions of how productively ecient and envi- deterministic and stochastic variants of a translog
ronmentally ecient Dutch dairy farming is with output distance function, and to obtain estimates
respect to the major environmentally detrimental of productive eciency and the shadow prices of
variables, and whether each type of eciency has undesirable outputs, in the Finnish pulp and paper
improved or deteriorated since the ®rst regulations industry. Reinhard et al. (1999) estimated a sto-
became eective. To answer these questions a chastic translog production frontier, using a panel
comprehensive environmental eciency measure of Dutch dairy farms, in which nitrogen surplus
must be developed and computed appropriately. was treated as an environmentally detrimental in-
A variety of environmental performance in- put. They calculated technical eciency and envi-
dexes have been proposed in the past, based on ronmental eciency, the latter being de®ned as the
adjustments of conventional measures of produc- ratio of minimum feasible to observed use of ni-
tive eciency. The indexes can be categorised as trogen surplus, conditional on observed levels of
those which are calculated using deterministic the desirable output and the remaining inputs.
techniques, which can be either parametric or This paper is along two lines an extension of the
nonparametric, and those which are estimated approach developed by Reinhard et al. (1999).
using stochastic techniques, which are exclusively First, we extend their approach to multiple envi-
parametric. The indexes can also be categorised on ronmentally detrimental inputs. Second, we im-
the basis of whether they treat the environmental plement this approach using both Stochastic
eects as inputs or outputs. Frontier Analysis (SFA) and DEA, and we com-
Fare et al. (1989) treated environmental eects pare the two sets of ®ndings.
as undesirable outputs, and they developed a hy- We de®ne environmental eciency as the ratio
perbolic eciency measure that evaluates producer of minimum feasible to observed use of environ-
performance in terms of the ability to obtain an mentally detrimental inputs, conditional on ob-
equiproportionate increase in desirable outputs served levels of the desirable output and
and reduction in undesirable outputs. They de- conventional inputs. This measure allows for a
veloped their measure on a strongly disposable radial reduction of the environmentally detrimen-
technology (applicable if undesirable outputs are tal inputs applied. Our measure distinguishes from
freely disposable) and on a weakly disposable the above mentioned environmental eciency in-
technology (applicable when it is costly to dispose dexes (except Reinhard et al., 1999) in the sense
o undesirable outputs, perhaps due to regulatory that we treat the environmentally detrimental
action). They proposed a nonparametric mathe- variables as inputs. Cropper and Oates (1992) and
matical programming technique known as Data Boggs (1997) also followed this approach. They
Envelopment Analysis (DEA) to construct strong- speci®ed a production function to include a vector
disposal and weak-disposal best-practice produc- of conventional inputs and the quantity of waste
tion frontiers, and to calculate their hyperbolic discharges. Waste emissions are treated simply as
eciency measure. F are et al. (1993) also treated another factor of production. Reductions in these
environmental eects as undesirable outputs, and emissions result in reduced output. Pittman (1981)
used a parametric mathematical programming also modelled pollution as an input in the pro-
technique similar to goal programming to calcu- duction function because the relation between an
late the parameters of a deterministic translog environmentally detrimental variable and output
output distance function. This enabled them to behaves like the relation between a conventional
calculate a hyperbolic eciency measure, and also input and output.
S. Reinhard et al. / European Journal of Operational Research 121 (2000) 287±303 289
We are able to measure the environmentally of variables increases. We will analyse the
detrimental input usage (excess nitrogen and ex- strengths and weaknesses of the two methods in
cess phosphate application and total energy use), computing the comprehensive environmental e-
and we know that they have undesirable environ- ciency scores.
mental repercussions, but we are unable to mea- The paper is organised as follows. In Section 2
sure the repercussions, and so we cannot the production process of dairy farms, including
incorporate undesirable outputs into our analysis. the environmentally detrimental inputs, is de-
The advantage of this comprehensive environ- scribed to provide the variables that have to be
mental eciency measure compared to the tradi- modelled. The concepts of technical and environ-
tional single input- single output measures, like mental eciency are elaborated in Section 3. In
excess nitrogen per hectare or energy use per ki- Section 4 the environmental eciency of each farm
logram of milk, becomes apparent when more is modelled within the context of a stochastic
than one environmentally detrimental input is in- translog production frontier. The corresponding
volved. DEA model is given in Section 5. The data are
There are two essential dierences between the described in Section 6; they summarise the pro-
econometric approach and mathematical pro- duction activities of an unbalanced panel of 613
gramming methods to the construction of a pro- Dutch dairy farms over the period 1991±1994.
duction frontier and the calculation of eciency Farm-level technical and environmental ecien-
relative to the frontier. The econometric approach cies of the two models are calculated, evaluated
has the virtue of being stochastic, and so attempts and compared in Section 7. Strengths and weak-
to distinguish the eects of statistical noise from nesses of the two models are treated in Section 8.
those of productive ineciency. However the Conclusions are formulated in Section 9.
econometric approach is parametric, and so can
confound the eects of misspeci®cation of (even
¯exible) functional forms (of both technology and 2. The Dutch dairy sector and the environment
ineciency) with ineciency. In addition, a ¯exi- problem
ble form is susceptible to multicollinearity, and
theoretical restrictions may be violated. A main Milk production takes place on about 37,400
attraction of the econometric approach is the farms in The Netherlands. The majority (80%) of
possibility it oers for a speci®cation in the case of these farms specialise in dairy farming. The dairy
panel data. It also allows for a formal statistical sector has a rather intensive character, although
testing of hypotheses and the construction of the total number of cows has decreased since the
con®dence intervals (Hjalmarsson et al., 1996). implementation of a milk quota system in 1984. In
Coelli (1995) concludes that the stochastic frontier 1995 1.7 million dairy cows were kept, and the
method is recommended for use in agricultural average specialised dairy farm maintained about
applications, because measurement error, missing 49 cows on 29 ha.
variables and weather, etc. are likely to play a Nitrogen and phosphate surplus and energy use
signi®cant role in agriculture. The mathematical are the three main environmental problems caused
programming approach is nonstochastic and by dairy farms. Nitrogen and phosphate surplus is
lumps noise and ineciency together and calls the induced by excess application of manure and
combination ineciency. The DEA version of the chemical fertiliser. Part of these nutrients are taken
mathematical programming approach is non- up by plants, but a large part is emitted to the
parametric, and less prone than SFA to speci®ca- environment. Nitrogen pollution leads to nitrate
tion error. It also imposes regularity conditions a contamination of the groundwater aquifers, the
priori rather than testing them ex post. DEA has most important source of drinking water. Nitro-
the additional advantage of being able to accom- gen also evaporates as ammonia and causes acid
modate many inputs and many outputs, although rain. Phosphate pollution causes eutrophication of
it generates more ecient ®rms when the number surface water, which endangers plant and ®sh life.
290 S. Reinhard et al. / European Journal of Operational Research 121 (2000) 287±303
The dairy sector is the second largest energy using roughage). Total energy use measures the appli-
sector in Dutch agriculture. Fossil energy use cation of energy, and is a proxy for the emission of
contributes to global warming, due to the emission CO2 , which contributes to global warming.
of carbon dioxide. A schematic representation of the main ¯ows of
To deal with these problems the Dutch have the environmentally detrimental variables is given
implemented a three-phase National Environ- in Fig. 1. Variables that aect the nitrogen and
mental Policy Plan (NEPP). One of the objectives phosphorus cycle directly, and variables that
of this plan is to achieve manuring that is balanced contain energy, are presented with the average
with the extraction of nutrients. Among other quantity of the corresponding environmentally
things, it has established increasingly restrictive detrimental input between brackets. The produc-
farm manure quotas. These quotas are ®rst based tion process on a dairy farm consists of two parts:
on the phosphate content of manure, because (i) roughage production that provides intermediate
phosphate does not evaporate, and is therefore input (grass and green maize) for the livestock and
more easily controlled. Fees are levied on phos- (ii) animal production that produces marketable
phate surpluses, and restrictions are imposed on outputs and manure. The latter is an intermediate
the spreading of manure. To implement the ma- output of the animal production process that is
nure policy with respect to nitrogen, nitrogen in- used in the roughage production. Both processes
puts and outputs are monitored by means of are depicted in Fig. 1. The inputs (including the
nutrient balance sheets. These in turn permit an intermediate input) are located left and the outputs
accurate calculation of farm-level nitrogen surplus, (including the intermediate output) are on the
the dierence between the quantity of nitrogen right.
applied and the quantity of nitrogen in the desir- The average nutrient input per farm in the data
able output. These balance sheets also provide set is 17,750 kg N and 1746 kg P (excluding in-
accurate computation of the phosphate surplus. A termediate input). The average nutrient output per
surplus measures the emission of minerals into the farm contains 17,750 kg N and 1746 kg P (in-
environment. While the environmental eects cluding the nitrogen and phosphate surplus). The
themselves are dicult to quantify, the mineral marketable output (milk, meat, livestock and
surpluses which create these eects can be quan- roughage) contains 3122 kg N and 592 kg P on
ti®ed. average. Thus the average nutrient surplus (nutri-
A goal of the Dutch government is an energy ent exchange with the soil and evaporation of
productivity increase of 26% by 2000 compared ammonia from land) consists of 82.4% of the N
with 1989. To this end the direct energy use of input and 66.1% of the P input. The total energy
dairy farms (and other small scale users) is taxed use contains 2,581 gigajoule, with the indirect en-
with an environment levy. To design and evaluate ergy use being far more important (88%) than di-
this energy policy the energy use of farms is rect energy use.
monitored. Dairy farms use fossil energy directly
as diesel and gas for heating, and they use elec-
tricity for milking machines and refrigeration of 3. Environmental eciency in the multiple environ-
milk. Dairy farms also apply inputs that contain mentally detrimental input case
fossil energy at an earlier stage in the product
chain. This energy use is referred to as indirect Environmental eciency is de®ned as the ratio
energy. For instance a lot of energy is used to of minimum feasible to observed use of environ-
produce nitrogen fertiliser. Also concentrates mentally detrimental inputs, conditional on ob-
contain an implicit amount of fossil energy. We served levels of output and the conventional
take the direct and indirect use of energy into ac- inputs. So de®ned, environmental eciency is a
count, to prevent intensive dairy farms (which buy nonradial input-oriented measure of technical ef-
concentrates) from unfairly appearing more ener- ®ciency that allows for a radial reduction of the
gy ecient than extensive farms (which grow environmentally detrimental inputs. Details are
S. Reinhard et al. / European Journal of Operational Research 121 (2000) 287±303 291
Fig. 1. Speci®cation of the average nitrogen, phosphate and energy ¯ows of dairy farms in the data set.
provided by Kopp (1981) and Banker and Morey environmentally detrimental inputs ZK and ZM ,
(1986). Fig. 2 portrays a production frontier in conditional on observed values of output Y and
conventional input X and environmentally detri- conventional input X, and Y 6 F
X ; ZK ; ZM . The
mental input Z space, holding output constant at comprehensive environmental eciency measure is
its observed value YR . The environmental e- calculated as a radial contraction of the two en-
ciency in this one bad input case is equal to vironmentally detrimental inputs to the frontier.
jOZ G j=jOZR j. The comprehensive environmental eciency index
In Fig. 3 the de®nition of environmental e- EER is de®ned as
ciency is extended to two environmentally detri-
mental inputs. This method can also be applied to jOZKF j jOZMF j
EER
more than two environmentally detrimental in- jOZKR j jOZMR j
puts. The relation between the two bad inputs is minfh : F
XR ; hZKR ; hZMR P YR g;
1
assumed to be similar to the relation between a
conventional input and a bad input as drawn in where ZKF hZKR ; ZMF hZMR is the minimum fea-
Fig. 2. Fig. 3 portrays a production frontier with sible environmentally detrimental input use, given
292 S. Reinhard et al. / European Journal of Operational Research 121 (2000) 287±303
F
and the observed conventional inputs XR and depicted in Fig. 3; the environmental eciency
output YR . Fare and Lovell (1978) show that the based solely on ZK is equal to jOZKG j=jOZKR j. This is
radial eciency measure is greater than or equal to smaller than the comprehensive environmental
the corresponding nonradial measure. This is also eciency measure, which equals jOZKF j=jOZKR j.
S. Reinhard et al. / European Journal of Operational Research 121 (2000) 287±303 293
In Fig. 2 the observed output is technically in- independently and identically distributed as
ecient, since R lies above the best practice pro- N(0,rv 2 ), intended to capture the in¯uence of ex-
duction frontier F
. It is possible to measure ogenous events beyond the control of farmers; Ui ,
technical eciency using an input-conserving ori- is a nonnegative random error term, independently
entation, as the ratio of minimum feasible input and identically distributed as N (l,ru 2 ).
use to observed input use, conditional on tech- The stochastic version of the output-oriented
nology and observed output production. It is also technical eciency measure (2) is given by the
possible to measure technical eciency using an expression
output-expanding orientation, as the ratio of ob-
Yit
served to maximum feasible output, conditional on TEit exp
ÿUi ;
4
technology and observed input usage. As F are and f
Xit ; Zit ; b; c; f exp
Vit
Lovell (1978) have noted, only under constant re-
Since Ui P 0; 0 6 exp
ÿUi 6 1. In order to im-
turns to scale do the two measures coincide for a
plement (4), technical ineciency must be sepa-
technically inecient producer. Not wishing to
rated from statistical noise in the composed error
impose constant returns to scale on the structure
term (Vit ÿ Ui ). Battese and Coelli (1988, 1992)
of production technology, we need to select an
have proposed the technical eciency estimator
orientation. We think an output orientation is
more appropriate in the current context, and so TEit E expfÿUi gj
Vit ÿ Ui :
5
our measure of technical eciency TER is given by
To derive a stochastic version of the environ-
ÿ1
TER maxf/ : /YR 6 F
XR ; ZR g mental eciency measure (1) we need to specify a
functional form for the deterministic kernel of the
j0YR j=0Y F ;
2
stochastic production frontier. To avoid excessive
misspeci®cation we use a ¯exible translog func-
where YF is maximum feasible output.
tional form to model the production technology.
(For convenience the farm and time subscripts on
the inputs X and Z are suppressed.) Writing (3) in
4. Stochastic frontier analysis translog form gives
year superscript t. This production set is convex tion, and no systematic errors due to nonresponse
and allows for variable returns to scale. Environ- are found (Dijk, 1990). In a few observations the
mental eciency as de®ned in Section 3 can be phosphate surplus is negative. These observations
computed with a linear program which measures cause both theoretical and technical problems.
performance in terms of the ability of a producer These observations cannot readily be used in either
to contract its environmentally detrimental inputs, SFA or DEA, because the natural log of negative
given its output and its conventional inputs. The values is not de®ned, and because input-oriented
linear program is similar to the fourth model of DEA is not translation-invariant. In the optimal
Ball et al. (1994), and is expressed as situation the phosphorus input equals the phos-
phorus output. Hence negative values are not op-
max h timal and the soil will lose its fertility. Because
h;k
only a very small percentage of the observations
s:t:
(less than 1%) suers from a phosphate de®cit we
X
I
have decided to omit these observations.
y ot 6 kti yit ;
i1
We have a total of 1535 observations in this
X
I unbalanced panel, and so each farm appears on
xot kti xtij ;
12 average 2.5 times. The period 1991±1994 is chosen,
j P j 1; . . . ; J
i1 because detailed information describing the nitro-
X
I gen ¯ows at each farm is available from 1991 on-
hÿ1 zot
k P kti ztik ; k 1; . . . ; K wards. The inputs and the output we specify are
i1 based upon the production process of dairy farms.
X
I
The production process, including the nitrogen
kti P 0; kti 1:
¯ows, is depicted in Fig. 1.
i1
In the speci®cation we have chosen the con-
Output-oriented technical eciency is comput- ventional inputs are aggregated into three catego-
ed with a model in which eciency is re¯ected as ries (labour, capital and variable inputs), and the
the ability of a producer to expand output, given outputs are aggregated into a single index of dairy
conventional inputs and environmentally detri- farm output. Ball et al. (1994) used these variables
mental inputs. This model is comparable to the also, although they distinguished separate output
second model of Ball et al. (1994). Input-oriented indexes for animal and roughage production. If
technical eciency is computed as the ability of prices at the farm level are available in the FADN,
a producer to contract both conventional and they are used to calculate price indexes. If prices
environmentally detrimental inputs equipropor- are not present in the FADN, price indexes are
tionately, conditional on output (these models are borrowed from CBS/LEI-DLO (1996). The
the original DEA formulations of Banker et al. FADN contains information on the quantity of
(1984), and are described in Appendix B). milk produced and the value of sales to the milk
factory and to other customers. The price that
farmers receive from the factory depends on the
6. Data protein and fat content of the milk, and so milk
prices re¯ect dierences in quality. Some farmers
In this study we utilise data describing the sell home-made cheese and butter, or sell milk
production activities of 613 highly specialised directly to customers. If we should use an index of
dairy farms that were in the Dutch Farm Ac- the quantity of milk produced, the dierences in
countancy Data Network (FADN) for part or all prices between farms result from dierences in the
of the 1991±1994 period. The FADN is a strati®ed quality of outputs and from dierences in the
random sample. Strati®cation is based on farm composition of the components. Then this price
size, age of the farmer, region and type of farm. index becomes an endogenous variable. Therefore
The FADN represents 99% of the milk produc- we prefer an implicit quantity index. Implicit
296 S. Reinhard et al. / European Journal of Operational Research 121 (2000) 287±303
quantity indexes are obtained as the ratio of value contained in desirable outputs, is measured in ki-
to the price index and therefore output is in prices lograms. The phosphate surplus is calculated ac-
of a speci®c year, 1991 being the base year. The cordingly. Total energy is the summation of the
price index used in this study is the average of a (implicit) energy content of all inputs in the pro-
multilateral Tornqvist price index across farms for duction process. The characteristics of the data set
each year (Higgins, 1986; Caves et al., 1982). This are summarised in Table 1.
price index varies over years but not across farms,
implying that dierences in the composition of a
netput or quality are re¯ected in quantity (Cox 7. SFA and DEA results
and Wohlgenant, 1986). The same method is used
to aggregate capital and the variable inputs. The 7.1. Stochastic frontier analysis
output quantity index contains milk, meat, live-
stock and roughage sold. These all contain nutri- First the full translog stochastic production
ents, which are depicted in Fig. 1. Labour input frontier with all three environmentally detrimental
consists of family labour, measured in hours. inputs and a normal- truncated normal error dis-
Buildings, equipment, livestock (for breeding and tribution was estimated. This model was tested
utilisation) and land are the components of capital against simpli®ed models. The production elastic-
stock. The capital price index is calculated from ities with respect to phosphate range from ÿ0.040
the revaluations of the capital stock. The value of to 0.095 with a mean of ÿ0.003. The monotonicity
many components of capital stock (buildings, assumption for phosphate surplus was violated in
equipment and livestock) is known at the start- more than half of the observations. Because these
balance and end-balance of each year. The dier- parameter estimates also generated environmental
ence between the start-balance of year t and the eciency scores, which could not be interpreted we
end-balance of year tÿ1 is due to revaluation of tested also a model without phosphate surplus.
capital stock. The price of land is computed as a The full translog frontier without phosphate sur-
price index of land for the distinguished soil types. plus could not be rejected. Thus we decided to
Labour and capital are not represented in Fig. 1, delete the phosphate surplus variables in the sto-
because these inputs do not contain nutrients or chastic frontier approach. The remaining full
energy (apart from the livestock component of translog nitrogen surplus ± total energy model was
capital stock). The variable input quantity index tested against simpli®ed models. The tests were
contains hired labour, concentrates, roughage, fe- performed nested, so if a simpli®ed model was not
rtiliser and other variable inputs. Fertiliser, con- rejected the next tests were performed against the
centrates and roughage purchased are depicted in simpli®ed model. On the basis of these tests, pre-
Fig. 1. The nitrogen surplus is represented in Fig. 1 sented in Table 2, we select the Model 2 speci®-
as the sum of `nutrient exchange with the soil' and cation. (We also explored the option of including a
`ammonia from land'. The nitrogen surplus, the time trend instead of year dummies. The time
dierence between nitrogen input and nitrogen trend turned out to be negative for more than 40%
Table 1
Characteristics of the data set (1535 observations)
Variables Unit Mean Min Max Std. dev.
Output 1000 Õ91 NLG 400 56 1456 232
Labour Hours 4107 1100 11,050 1535
Capital 1000 Õ91 NLG 2259 431 8166 1143
Variable input 1000 Õ91 NLG 147 16 665 92
Nitrogen surplus kg N 14,628 1927 63,779 8764
Phosphorus surplus kg P 1154 2 8934 942
Energy Gigajoule 2581 321 24,213 1628
S. Reinhard et al. / European Journal of Operational Research 121 (2000) 287±303 297
Table 2
Speci®cation tests of the nitrogen surplus and total energy model for alternative stochastic frontier speci®cations
Speci®cation Null hypothesis Tested Loglike- Likelihood v2 Critical Decision
against lihood ratio value
1. Full translog 1633.063
2. Model 1 bjl 0; 8j 8l ; j 6 l 1 1633.063 0 7.82 Accepteda
3. Model 2 fjk 0; 8j 8k 2 1629.517 7.09 16.92 Accepteda
4. Half-normal l0 3 1149.898 959.24 3.84 Rejected
a
This model is the new basis (null) model.
of all observations. We cannot ®nd an economic nitrogen surplus is small, and the monotonicity
explanation for this result. The time trend variable assumption is violated for nitrogen surplus in 21%
seems to capture something other than techno- of the observations. The mean output elasticity of
logical change, for instance regulation.) The se- energy is very large, even larger than the mean
lected model contains a truncated normal output elasticity of variable inputs, possibly due to
distribution of the nonnegative random term U. multicollinearity with these variable inputs. The
The parameter estimates of the translog frontier means of the elasticities hardly change through
are presented in Appendix A. This model is used to time.
generate the technical eciency scores and the We now turn to the estimated technical and
parameter estimates required to compute envi- environmental eciencies, which are summarised
ronmental eciency scores. in Table 4. Output-oriented technical eciency is
Before turning to an investigation of technical estimated using (5). Input-oriented technical e-
and environmental eciency, we ®rst consider the ciency is estimated as the maximum radial con-
structure of the estimated production technology. traction of all inputs (conventional and
Table 3 reports elasticities of output with respect environmentally detrimental); in the same way as
to each input. The sum of the elasticities of output in Reinhard et al. (1999). Due to the presence of
with respect to the ®ve inputs generates an esti- increasing returns to scale, input-oriented techni-
mated scale elasticity which indicates the presence cal eciency scores are higher than output-ori-
of increasing returns to scale, a ®nding that is ented technical eciency scores at almost all
qualitatively comparable to, but somewhat smaller observations.
than, that of Reinhard et al. (1999). The elasticities The output-oriented and input-oriented tech-
of output with respect to each of the conventional nical eciency scores seem very plausible. As ex-
inputs are positive for 100% of the observations pected, the environmental eciency scores are
and are also in line with the ®ndings of Reinhard smaller, although not dramatically so. Also as
et al. (1999). The mean elasticity with respect to expected, the nonradial environmental eciency
Table 3
Elasticities of output in SFA
Labour Capital Variable Nitrogen Energy Total
inputs
Overall mean 0.079 0.314 0.212 0.021 0.464 1.090
Lower quartile 0.072 0.292 0.194 0.003 0.415 1.051
Median 0.080 0.312 0.215 0.020 0.462 1.088
Upper quartile 0.087 0.333 0.231 0.037 0.512 1.131
Mean 1991 0.079 0.313 0.211 0.019 0.470 1.093
1992 0.079 0.313 0.212 0.021 0.464 1.090
1993 0.079 0.314 0.213 0.021 0.462 1.089
1994 0.079 0.314 0.214 0.023 0.459 1.090
298 S. Reinhard et al. / European Journal of Operational Research 121 (2000) 287±303
Table 4
SFA technical and environmental eciency scores
Technical (%) Environmental (%)
Output Input All bads Nitrogen Energy
Overall mean 88.94 89.88 79.54 27.20 79.32
Lower quartile 84.57 85.78 70.76 13.91 70.71
Median 90.73 91.55 81.85 23.78 81.27
Upper quartile 94.85 95.29 89.68 36.50 89.39
Mean 1991 89.01 89.96 79.80 26.83 79.66
1992 88.77 89.72 79.31 27.13 79.11
1993 88.97 89.89 79.51 27.26 79.27
1994 89.03 89.95 79.55 27.57 79.27
scores are smaller than the radial input-oriented year to year. Input-oriented technical eciency,
technical eciency scores. To analyse the envi- comprehensive environmental eciency and ener-
ronmental eciency scores more closely, the en- gy eciency also show no trend. Nitrogen e-
vironmental eciency with respect to the two bad ciency increases slightly through time, suggesting
inputs separately is also presented. The nitrogen that the regulatory impact has been small but
eciency scores are very low. Nitrogen is applied positive.
ineciently because in dairy farming because the
nitrogen surplus is hardly sanctioned yet. The en-
ergy eciency scores are marginally lower than the
comprehensive environmental eciency scores. 7.2. Data envelopment analysis
Thus energy eciency determines to a very large
extent the value of the environmental eciency The DEA results are based on all three envi-
score. ronmentally detrimental inputs, since unlike SFA,
Output-oriented technical eciency is assumed phosphate causes no diculties in DEA. The
to be constant through time during the research output elasticities calculated from the dual to the
period. The time-invariant speci®cation is not un- DEA program (12) are reported in Table 5, and
reasonable, since at most four observations per are generally in line with the corresponding SFA
farm, and on average 2.5 observations per farm, results. The mean nitrogen elasticity is again pos-
are available in the data set. However due to the itive and small, as is the mean phosphate elasticity.
unbalanced nature of the panel, the output-ori- The mean energy elasticity is again larger than the
ented technical eciency scores dier slightly from mean variable input elasticity. The mean returns to
Table 5
Elasticities of output in DEA
Labour Capital Variable Nitrogen Phosphate Energy Total
inputs
Overall mean 0.159 0.295 0.204 0.072 0.022 0.359 1.112
Lower quartile 0.000 0.126 0.055 0.000 0.000 0.254 0.963
Median 0.093 0.315 0.148 0.000 0.006 0.369 1.058
Upper quartile 0.236 0.439 0.322 0.109 0.027 0.484 1.178
Table 6
DEA Technical and environmental eciency scores
Technical (%) Environmental (%)
Output Input All bads Nitrogen Phosphate Energy
Overall mean 78.37 81.10 51.95 40.82 18.98 53.30
Lower quartile 70.10 73.74 33.84 24.07 2.87 37.71
Median 77.54 80.00 50.73 35.09 6.57 52.67
Upper quartile 86.35 88.15 65.49 50.53 19.63 64.53
Mean 1991 78.54 81.27 52.57 42.17 20.50 54.63
1992 78.21 80.91 51.85 40.54 18.67 53.05
1993 77.64 80.44 50.74 39.25 17.60 51.67
1994 79.12 81.84 52.77 41.53 19.34 54.05
scale is marginally larger than the SFA estimate. SFA score. The absence of meaningful trends in
All in all, there is a high degree of concordance the DEA technical and environmental eciency
between the SFA and DEA characterisations of scores has no implication for trends in absolute
production technology. performance, since separate production sets have
The DEA models described in Section 5 are been constructed for each year. The only implica-
used to compute the eciency scores presented in tion which can be drawn from the absence of
Table 6. The output-oriented and input-oriented trends in technical and environmental eciency
eciency scores are on average somewhat lower scores, is that dispersion within each annual sam-
than the corresponding SFA eciency scores, ple has remained virtually unchanged.
presumably because SFA incorporates the eects Table 7 reports rank correlations between the
of random noise. The comprehensive environ- SFA scores and the corresponding DEA scores.
mental eciency scores (based on the three envi- All rank correlations are positive, ranging from
ronmentally detrimental inputs) and energy 0.76 for output-oriented technical eciency to 0.46
eciency scores are also smaller than the corre- for nitrogen eciency. The rank correlation coef-
sponding SFA eciency scores, presumably for ®cient for the comprehensive environmental e-
the same reason. However in DEA the mean en- ciency scores is 0.49. The rank correlations
ergy eciency score is larger than the mean envi- between the SFA output-oriented technical e-
ronmental eciency score. The mean nitrogen ciency scores and the other SFA eciency scores
eciency score is larger than the corresponding are quite high, ranging from 1.0 for input-oriented
Table 7
Rank correlation coecients of the eciency measuresa
TE-O TE-I EE-total EE-N EE-E
SFA DEA SFA DEA SFA DEA SFA DEA SFA DEA
TE-O-SFA 1.0 0.76 1.0 0.69 0.99 0.50 0.88 0.43 0.98 0.52
TE-O-DEA 1.0 0.76 0.90 0.74 0.72 0.74 0.70 0.71 0.71
TE-I-SFA 1.0 0.70 1.0 0.50 0.86 0.43 0.99 0.52
TE-I-DEA 1.0 0.70 0.72 0.61 0.72 0.68 0.71
EE-SFA 1.0 0.49 0.82 0.42 0.99 0.51
EE-DEA 1.0 0.50 0.79 0.47 0.92
EE-N-SFA 1.0 0.46 0.78 0.53
EE-N-DEA 1.0 0.40 0.76
EE-E-SFA 1.0 0.49
EE-E-DEA 1.0
a
O output, I input, N nitrogen surplus and E energy.
300 S. Reinhard et al. / European Journal of Operational Research 121 (2000) 287±303
technical eciency to 0.88 for nitrogen eciency. technique which does not. Statistical noise ac-
The rank correlation between the DEA output- counts for nearly 30% of the SFA regression re-
oriented technical eciency scores and the other siduals (Appendix A). This accounts for the fact
DEA eciency scores is smaller, ranging from 0.90 that SFA eciency scores are generally higher
for input-oriented technical eciency to 0.70 for than DEA eciency scores, despite the fact that
nitrogen eciency. the characteristics of the two production frontiers
are so similar. While eorts are underway to make
DEA stochastic (e.g., Land et al., 1993; Olesen
8. Strengths and weaknesses of SFA and DEA with and Petersen, 1995), these eorts are in their in-
respect to environmental eciency fancy.
The ability to impose or test theoretical re-
SFA and DEA have constructed remarkably strictions: DEA satis®es monotonicity and cur-
similar production frontiers from the data. A vature restrictions by construction, so the absence
comparison of Table 3 and Table 5 con®rms this of a testing procedure is inconsequential. How-
concordance. Both techniques suggest the presence ever these restrictions cannot be imposed in SFA
of scale economies of the order of 1.10. Both when a ¯exible translog functional form is spec-
techniques provide the same ranking of output i®ed, which makes the ability to test for the sat-
elasticities, ranging from energy, capital, variable isfaction of the theoretical restrictions a strength
inputs and labour down to nitrogen, although the of SFA. The SFA model with all three environ-
magnitudes vary in some cases. This concordance mentally detrimental inputs included did not
is reassuring. make sense. We encountered occasional viola-
However a comparison of Tables 4 and 6 sug- tions of monotonicity in the SFA model with two
gests that SFA and DEA generate somewhat less of three environmentally detrimental inputs in-
similar technical and environmental eciencies cluded. When we tested for monotonicity (but
relative to their similar production frontiers. SFA not for curvature), the test uncovered failure for
technical eciency scores are higher (by about 21% of the observations for the nitrogen surplus
10%) than DEA eciency scores, and exhibit less input.
variability. SFA comprehensive environmental ef-
®ciency scores and energy eciency scores are also
higher (by nearly 30%) than comparable DEA
scores, while SFA nitrogen eciency scores are 9. Conclusions
considerably lower than comparable DEA scores.
Although SFA and DEA generate somewhat dif- In this paper we have developed an analytical
ferent mean eciency score magnitudes, an in- framework within which to calculate environmen-
spection of Table 7 con®rms that the two tal eciency in the presence of multiple environ-
techniques generate similar rankings of individual mentally detrimental inputs. This comprehensive
farms on the basis of the various eciency criteria. environmental eciency measure can identify
This is also reassuring. farms with the smallest and the largest environ-
We now consider strengths and weaknesses of mentally detrimental emissions to the environ-
the two techniques. In contrast to other studies in ment, in relation to their use of conventional
which SFA and DEA have been compared (e.g., inputs and the output they produce. Our measure
Ferrier and Lovell (1990), Hjalmarsson et al. enables the aggregation of environmentally detri-
(1996)), we consider evaluation criteria which are mental inputs, and it allows the calculation of the
speci®c to our objective, the measurement of environmental eciency of the distinguished en-
technical and environmental eciency. vironmentally detrimental inputs. It also indicates
The ability to account for exogenous in¯uences: which environmentally detrimental input is used
SFA is a stochastic technique which contains a most ineciently, both on individual farms and in
random error term, while DEA is a deterministic the aggregate.
S. Reinhard et al. / European Journal of Operational Research 121 (2000) 287±303 301
We have found Dutch dairy farms to have DEA is deterministic, and is unable to identify
respectable levels of technical eciency (78±89% whether the environmentally detrimental variables
on average, depending on the empirical tech- suit the model.
nique). We have also found the comprehensive
environmental eciency score to be somewhat
smaller on average (52±80% on average, depend- Appendix A
ing on the empirical technique). We have found
energy to be utilised more eciently than nitro- See Table 8.
gen, with mean nitrogen eciency scores ranging
from 27% to 41%, depending on the empirical
technique). Energy eciency determines to a very Appendix B. DEA models for output-oriented and
large extent the value of the environmental e- input-oriented technical eciency
ciency score.
We have tested this framework extensively with The linear programming model to compute
two methods, SFA and DEA. We have evaluated output-oriented technical eciency
these methods on the basis of their ability to in-
corporate the eects of statistical noise, their ful- max h
h;k
®lment of theoretical restrictions, and the s:t:
possibility to test these restrictions. Both SFA and
DEA can estimate environmental eciency scores, X
I
hy ot 6 kti yit ;
although only SFA incorporates noise. However i1
SFA allows the estimation of environmental e- X
I
ciency scores only in the two environmentally xot
j P kti xtij ; j 1; . . . ; J
detrimental input case. The three bad input case i1
(including phosphate surplus) did not ful®l all X
I
Table 8
Parameter estimatesa
Parameter Coecient estimate Standard error Parameter Coecient Standard error
estimate
b0 ÿ4.500 1.903 cNN ÿ0.068 0.037
b1 0.342 0.325 cEE ÿ0.201 0.017
bc 1.073 0.318 cNE 0.094 0.029
bv ÿ0.267 0.267 bD92 ÿ0.014 0.009
cN ÿ0.067 0.174 bD93 ÿ0.019 0.009
cE 1.124 0.223 bD94 0.004 0.009
bll ÿ0.032 0.039 l/ru 0.468 0.0002
bcc ÿ0.052 0.022 ru 2 /rv 2 2.452 0.255
bvv 0.041 0.023 rv 2 0.081 0.0004
a
The subscripts l,c,v,N and E refer to labour, capital, variable input, and nitrogen surplus and energy, respectively. D92,D93 and D94
refer to the yeardummies.
302 S. Reinhard et al. / European Journal of Operational Research 121 (2000) 287±303
max h Charnes, A., Clark, C.T., Cooper, W.W., Golany, B., 1985. A
h;k developmental study of data envelopment analysis in
s:t: measuring the eciency of maintenance units in the US
air forces. Annals of Operations Research 2, 95±112.
X
I
Coelli, T.J., 1995. Recent developments in frontier modelling
y ot 6 kti yit ; and eciency measurement. Australian Journal of Agricul-
i1 tural Economics 39 (3), 219±245.
X
I Cox, T.L., Wohlgenant, M.K., 1986. Prices and quality eects
hÿ1 xot
j P kti xtij ; j 1; . . . ; J ; in cross-sectional demand analysis. American Journal of
i1 Agricultural Economics 68, 908±919.
X
I Cropper, M.L., Oates, W.E., 1992. Environmental economics:
hÿ1 zot
k P kti zikt ; k 1; . . . ; K; A survey. Journal of Economic Literature 30 (2), 675±740.
i1 Dijk, J., 1990. Representativeness and precision of Dutch
farming sector income statistics. Paper presented at the
X
I
VI-th Congress of the European Association of Agricultural
kti P 0; kti 1:
Economists, The Hague.
i1
F
are, R., Grosskopf, S., Lindgren, B., Roos, P., 1992. Produc-
tivity changes in Swedish pharmacies 1980±1989: A non-
parametric Malmquist approach. Journal of Productivity
Analysis 3 (1/2), 85±102.
References F
are, R., Grosskopf, S., Lovell, C.A.K., Pasurka, C., 1989.
Multilateral productivity comparisons when some outputs
Aigner, D.J., Lovell, C.A.K., Schmidt, P., 1977. Formulation are undesirable: A nonparametric approach. The Review of
and estimation of stochastic frontier production function Economics and Statistics 71 (1), 90±98.
models. Journal of Econometrics 6 (1), 21±37. F
are, R., Grosskopf, S., Lovell, C.A.K., Yaisawarng, S., 1993.
Ball, V.E., Lovell, C.A.K., Nehring, R.F., Somwaru, A., 1994. Derivation of shadow prices for undesirable outputs: A
Incorporating undesirable outputs into models of produc- distance function approach. The Review of Economics and
tion: An application to US agriculture. Cahiers d'Economi- Statistics 75 (2), 374±380.
que et Sociologie Rurales 31, 59±73. F
are, R., Lovell, C.A.K., 1978. Measuring the technical
Banker, R.D., Charnes, A., Cooper, W.W., 1984. Some models eciency of production. Journal of Economic Theory 19
for estimating technical and scale ineciencies in data (1), 150±162.
envelopment analysis. Management Science 30 (9), 1078± Ferrier, G.D., Lovell, C.A.K., 1990. Measuring cost eciency
1092. in banking: Econometric and linear programming evidence.
Banker, R.D., Morey, R., 1986. Eciency analysis for exoge- Journal of Econometrics 46, 229±245.
nously ®xed inputs and outputs. Operations Research 34 Hetem aki, L., 1996. Essays on the impact of pollution control
(4), 513±521. on a ®rm: A distance function approach. Finnish Forest
Battese, G.E., Coelli, T.J., 1992. Frontier production functions, Research Institute Research Papers 609, Helsinki.
technical eciency and panel data: With application to Higgins, J., 1986. Input demand and output supply on Irish
paddy farmers in India. Journal of Productivity Analysis 3 farms ± a micro-economic approach. European Review of
(1/2), 153±169. Agricultural Economics 13, 477±493.
Battese, G.E., Coelli, T.J., 1988. Prediction of ®rm-level Hjalmarsson, L., Kumbhakar, S.C., Heshmati, A., 1996. DEA,
technical eciencies with a generalized frontier production DFA and SFA: A comparison. Journal of Productivity
function and panel data. Journal of Econometrics 38, 387± Analysis 7 (2/3), 303±328.
399. Judge, G., Hill, C., Griths, W., Lee, T., Lutkepol, H., 1982.
Belsley, D.A., Kuh, E., Welsch, R.E., 1980. Regression Diag- An Introduction to the Theory and Practice of Economet-
nostics: Identi®cation of In¯uential Data and Sources of rics. Wiley, New York.
Collinearity. Wiley, New York. Kopp, R.J., 1981. The measurement of productive eciency: A
Boggs, R.L., 1997. Hazardous waste treatment facilities: reconsideration. Quarterly Journal of Economics 96 (3),
Modeling production with pollution as both an input and 477±503.
an output. Unpublished Ph.D. Dissertation, University of Land, K, Lovell, C.A.K., Thore, S., 1993. Chance-constrained
North Carolina, Chapel Hill, NC 27599, USA. data envelopment analysis. Managerial and Decision Eco-
Caves, D.W., Christensen, L.R., Diewert, W.E., 1982. Multi- nomics 14 (6), 541±554.
lateral comparisons of output, input and productivity using Maddala, G.S., 1988. Introduction to Econometrics. Macmil-
superlative index numbers. The Economic Journal 92, 73± lan, New York.
86. Meeusen, W., van den Broeck, J., 1977. Eciency estimation
CBS/LEI-DLO, 1996. Agricultural Data 1996 (in Dutch). CBS/ from Cobb±Douglas production functions with composed
LEI-DLO, Voorburg/The Hague. error. International Economic Review 18 (2), 435±444.
S. Reinhard et al. / European Journal of Operational Research 121 (2000) 287±303 303
Olesen, O.B., Petersen, N.C., 1995. Chance constrained e- Stevenson, R.E., 1980. Likelihood functions for generalized
ciency evaluation. Management Science 41 (3), 442±457. stochastic frontier estimation. Journal of Econometrics 13
Pittman, R.W., 1981. Issues in pollution control: Interplant cost (1), 58±66.
dierences and economies of scale. Land Economics 57 (1), Tyteca, D., 1997. Linear programming models for the mea-
1±17. surement of environmental performance of ®rms ± concepts
Reinhard, S., Lovell, C.A.K., Thijssen, G., 1999. Econometric and empirical results. Journal of Productivity Analysis 8 (2),
estimation of technical and environmental eciency: An 183±198.
application to Dutch dairy farms. American Journal of
Agricultural Economics 81 (1), 44±60.