You are on page 1of 12

ANALYSIS OF ENVIRONMENTAL

EFFICIENCY VARIATION
STIJN REINHARD, C. A. KNOX LOVELL, AND GEERT THIJSSEN
In this article, we develop and implement a methodology for analyzing the sources of variation in
environmental efciency across producers. We formulate a two-stage model. In the rst stage, we use
stochastic frontier analysis toestimate bothtechnical andenvironmental efciency. Inthe secondstage,
we again use stochastic frontier analysis to regress estimated environmental efciency scores against a
variety of technology, physical environment, and management variables. In this stage we estimate the
impact of each explanatory variable on environmental efciency, and we derive conditional estimates
of environmental efciency from the one-sided error component. We illustrate our methodology with
an empirical application to a panel of Dutch dairy farms. We nd evidence of relatively low levels of
environmental efciency, and we nd that environmental efciency can be improved through a number
of policy options, including the provision of farmers with more insight into the nutrient balance of their
farms.
Key words: agriculture, environmental efciency, nitrogen surplus.
The empirical analysis of any type of efciency
should have two components: (a) the esti-
mation of its variation across producers, and
(b) the identication of its determinants. The
former provides an indication of the severity
of the inefciency problem, and the latter pro-
vides evidence on the sources of the problem.
The former has been widely studied, in agri-
culture and elsewhere, although the latter has
been studied less frequently.
1
Both have tend-
ed to examine the nature and sources of
technical efciency rather than environmental
efciency, and the distinction is important.
In this article, we develop a methodology for
an empirical analysis of the sources of vari-
StijnReinhardis a senior researcher at the Agricultural Economics
Research Institute (LEI), The Hague. Knox Lovell is a professor in
the Department of Economics at the University of Georgia. At the
time of this research, Geert Thijssen was associate professor in the
Department of Economics and Management at the Wageningen
University.
This research was carried out in cooperation with the Mansholt
Graduate School, Wageningen. The authors are grateful to Spiro
Stefanou and two AJAE reviewers for their suggestions and de-
tailed comments.
1
Farm management studies have investigated determinants of
farm success, typically measured in nancial terms, since the
early 20th century; Fox, Bergen and Dickson review these studies.
Second-stage regression analyses investigating the determinants
of farm productive efciency apparently began with Sitorus and
Timmer, who used Data Envelopment Analysis in the rst stage,
and Pitt and Lee, who used a xed effects regression in the rst
stage. We are not aware of any previous studies that have used
a second-stage regression to explain variation in environmental
efciency.
ation in environmental efciency. We study
Dutch dairy farming, where environmental
degradation has been severe, and where the
abatement of nitrogen emissions is a policy ob-
jective of the Dutch government. Recently, the
focus of Dutch policy analysis is on the effect
of various policies on environmental efciency
in Dutch agriculture.
Previous articles provided insight into how
environmentally efcient Dutch dairy farming
is (Reinhard, Lovell and Thijssen; Reinhard
andThijssen). There we foundevidence of sub-
stantial variation in environmental efciency
beneath best practice, which itself may be ab-
solutely inefcient. However, the formulation
of policy designed to improve the environmen-
tal performance of dairy farming requires that
the impact of various characteristics on envi-
ronmental efciency be identied. Therefore,
the objectives of this article are to identify
the sources of variation in environmental ef-
ciency across farms, to quantify their impacts,
and to reassess farm environmental efciency
in light of these characteristics. This agenda
raises two questions: (a) what variables are as-
sociated with variation in environmental ef-
ciency?; and (b) what methodology is most ap-
propriate to incorporate these variables into a
model of environmental efciency?
We proceed in two stages. In the rst stage,
we formulate and estimate a composed error
stochastic production frontier model, in which
Amer. J. Agr. Econ. 84(4) (November 2002): 10541065
Copyright 2002 American Agricultural Economics Association
Reinhard, Lovell, and Thijssen Analysis of Environmental Efficiency Variation 1055
conventional and environmentally detrimen-
tal inputs are combinedtoproduce marketable
output. In this framework estimates of techni-
cal efciency are extracted from the one-sided
error component, while estimates of environ-
mental efciency are derived from estimates
of parameters in the model, including both
technology parameters and parameters de-
scribing the distribution of the one-sided error
component.
In the second stage, we formulate and es-
timate a composed error stochastic environ-
mental efciency frontier model, in which the
environmental efciencies estimated in the
rst stage are regressed against a set of ex-
planatory variables. The explanatory variables
are identied in a comparison between a com-
prehensive dairy farming model
2
and the em-
pirical model we use to estimate (uncondi-
tional) environmental efciency.
Two types of information emerge from the
second stage. One is sample-wide evidence on
the directions and magnitudes of the impacts
of the explanatory variables onestimatedenvi-
ronmental efciency. This evidence is derived
from the estimated coefcients of the deter-
ministic part of the environmental efciency
frontier. The other is producer-specic evi-
dence on the ability of individual producers
to keep up with best practice environmental
efciency standards, conditional on their cir-
cumstances as characterized by their explana-
tory variables. This evidence is extracted from
the one-sided error component of the environ-
mental efciency frontier model.
Both types of information generated in the
second stage are useful for policy purposes.
The rst identies factors exerting signicant
impacts on the environmental efciency
of dairy farming. To the extent that these
factors are subject to government inuence,
they can be manipulated to improve overall
environmental performance. The second
identies factors that can be used in the
design of incentive-based regulation, since it
identies environmentally (in)efcient dairy
farms whose performance has been achieved
despite (or because of) these factors.
Our article is structured as follows. In the
next section we present our two-stage stochas-
tic frontier methodology, followed by a com-
prehensive model of dairy farming to identify
potential determinants of variation in environ-
2
Section Determinants of Environmental Efciency and Data
contains a description of this model, more details can be found in
Reinhard.
mental efciency, and a description of our data
set. The impacts of the explanatory variables
and the farm-level conditional environmental
efciency scores are then presented. The nal
section concludes with a summary and discus-
sion of our ndings.
Modeling, Estimating, and Explaining
Environmental Efciency
We dene environmental efciency EE as the
ratioof minimumfeasibletoobserveduseof an
environmentally detrimental input, giventech-
nology and the observed levels of output and
conventional inputs. Thus
EE = min{ : F(X, N) Y} 1 (1)
whereF() is theproductionfrontier, X R
N
+
is
a vector of conventional inputs, N R
+
is an
environmentally detrimental input,
3
and Y
R
+
is output.
Since output production typically takes
place after input decisions are made, technical
efciency TEis denedas the ratioof observed
to maximumfeasible output, given technology
and observed usage of all inputs. Thus
TE = [max{ : Y F(X, N)}]
1
1. (2)
Agricultural output is typically treated as a
stochastic variable because of weather condi-
tions, diseases, and other exogenous random
forces. We, therefore, convert the determinis-
tic relationship Y F(X, N) to the stochastic
production frontier
Y
i t
= F(X
i t
, N
i t
; ) exp{V
i t
U
i
} (3)
i = 1, . . . , I, t = 1, . . . , T
where F(X
i t
, N
i t
; ) is the deterministic ker-
nel of the stochastic production frontier
[F(X
i t
, N
i t
; ) exp{V
i t
}], is a technology
parameter vector to be estimated, V
i t

iid N(0,
2
v
) captures random events be-
yond the control of farmers, and U
i
i.i.d.
N
+
(,
2
u
) captures time-invariant technical
inefciency in production.
The stochastic versionof TE
i
is derivedfrom
(3) as
TE
i
= Y
i t
/[F(X
i t
, N
i t
; ) exp{V
i t
}] (4)
= exp{U
i
} 1, i = 1, . . . , I.
3
The environmentally detrimental input is labeled Nbecause in
our empirical workthis variableis surplus nitrogenemanatingfrom
use of fertilizer and manure; see Reinhard, Lovell and Thijssen for
more information on the environmental aspects of Dutch dairy
farming.
1056 November 2002 Amer. J. Agr. Econ.
The Battese and Coelli (1988) estimator for
TE
i
is
TE
i
= E[exp{U
i
}/(V
i t
U
i
)] (5)
=
_
1 (

i
/

)
1 (
i
/

)
_
exp
_

i
+
1
2

_
i = 1, . . . , I
where () is the standard normal distribution
function,

=
u

v
/(
2
u
+
2
v
)
1/2
, and
i
=
[(V
i t
U
i
)
2
u
+
2
v
]/(
2
u
+
2
v
). The tech-
nology and error component parameters
(,
2
v
,
2
u
, ) are estimated using maximum
likelihood techniques. Derived estimates of

and
i
are inserted into the second and third
line of (5) to generate estimates of TE
i
.
4
To derive a stochastic version of EEwe need
to specify a functional form for F(X
i t
, N
i t
; ).
Specifying (3) in translog form gives
ln Y
i t
=
0
+

j
ln X
i t j
+
n
ln N
i t
(6)
+
1
2

j k
ln X
i t j
ln X
i t k
+

j n
ln X
i t j
ln N
i t
+
1
2

nn
(ln N
i t
)
2
+ V
i t
U
i
.
where
j k
=
kj
. The logarithm of the output
of an environmentally efcient producer is ob-
tained by replacing N
i t
with N
i t
and setting
U
i
= 0. Setting (6) and the output of the envi-
ronmentally efcient producer equal and solv-
ing for lnEE
i t
= lnN
i t
ln N
i t
= ln yields
the environmental efciency estimator
ln EE
i t
(7)
=
_

n
+

j n
ln X
i t j
+
nn
ln N
i t
_

_
_

n
+

j n
ln X
i t j
+
nn
ln N
i t
_
2
2
nn
U
i
_
0.5
_
_
_

nn
.
4
It is alsopossibletoestimate(3) usingconventional xedeffects
or random effects panel data techniques, in which case estimates
of TE
i
would be obtained from normalized farm effects.
Environmental efciency is calculated using
the positive root in (7).
For the second-stage analysis of variation in
technical efciency, two methods have been
developed in literature; for an overview see
Kumbhakar and Lovell (chap. 7). The stan-
dard approach is to regress estimated tech-
nical efciencies against a set of explanatory
variables. OLS is frequently used (e.g., Hallam
and Machado), although a limited dependent
variable technique such as tobit is preferred
(e.g., Weersink, Turvey and Godah). However,
regardless of the estimation procedure em-
ployed, the two-stage approach suffers from
a fundamental inconsistency. It is assumed in
the rst stage that technical efciencies are in-
dependently identically distributed (iid), but
this assumption is contradicted in the second-
stage regression in which estimated technical
efciencies are assumed to have a functional
relationship with the explanatory variables.
Battese and Coelli (1995) originally raised
the objection to the use of estimates of TE
i
as
dependent variables in a second-stage regres-
sion. However, their argument does not ap-
ply to the similar use of estimates of EE
i t
. Al-
though technical efciency is estimated from
an error component, environmental efciency
is calculated from parameter estimates de-
scribing the structure of production technol-
ogy and the one-sided error component. This
is an important distinction. While the iid as-
sumption on the U
i
is inconsistent with the use
of estimates of TE
i
as dependent variables in
a second-stage regression, no such assumption
is made concerning EE
i t
. Thus it is permissible
touse estimates of EE
i t
as dependent variables
in a second-stage regression.
We depart from OLS or tobit convention by
using maximum likelihood techniques to esti-
mate a stochastic frontier regression model in
the second stage. Our strategy is to estimate
a stochastic environmental efciency frontier,
andtoobtainrevisedestimates of environmen-
tal efciency that are conditioned on variation
in the explanatory variables. A second-stage
frontier approach offers both economic and
statistical advantages over a second stage OLS
or tobit approach. The economic intuition be-
hind the frontier approach is that it provides
a characterization of the relationship between
best practice environmental efciency and the
explanatory variables, and it partitions devi-
ations from best practice to statistical noise
and conditional environmental inefciency re-
maining evenafter variationinthe explanatory
variables has been accounted for. Neither OLS
Reinhard, Lovell, and Thijssen Analysis of Environmental Efficiency Variation 1057
nor tobit can provide this information. From a
statistical perspective, our estimates of EE
i t
do
not approach, much less cluster at, their limit-
ing value of unity, and so tobit offers no advan-
tage over OLS.
5
In addition, OLS parameter
estimates are biased and inconsistent if the dis-
turbance term is skewed with nonzero mean.
6
Finally, the estimation of a stochastic frontier
enables us to test the symmetry and zero mean
hypotheses justifying OLS. The presence of
skewness can indicate that environmental in-
efciency remains even after accounting for
the impacts of the explanatory variables. Thus
estimation of a stochastic frontier generates
revised estimates of environmental efciency
conditionedonthevariationintheexplanatory
variables. Neither OLS nor tobit can provide
this information.
The stochastic environmental efciency
frontier regression model can be expressed in
general form as
EE
i t
= G(Z
i t
; ) exp
_
V

i t
U

i
_
(8)
i = 1, . . . , I, t = 1, . . . , T
where G(Z
i t
; ) is the deterministic kernel of
the stochastic environmental efciency fron-
tier [G(Z
i t
; ) exp{V

i t
}], Z
i t
is a vector of
explanatory variables expected to inuence
environmental efciency, is a vector of pa-
rameters to be estimated, V

i t
i.i.d. N(0,
2
v
)
and U

i
i.i.d. N
+
(

,
2
u
). In this formula-
tion variation in estimated environmental ef-
ciency is apportioned to three sources: (a) the
impacts of the explanatory variables captured
by G(Z
i t
; ); (b) the impact of measurement
error and other sources of statistical noise re-
ected in V

i t
; and (c) an unexplained shortfall
of environmental efciency beneath best prac-
tice observed in the sample reected in U

i
.
Conditional environmental efciency CEE
i t
is
dened and estimated exactly as in (4) and (5),
and so
CEE
i
= EE
i t
/
_
G(Z
i t
; ) exp
_
V

i t
__
(9)
= exp
_
U

i
_
1, i = 1, . . . , I
which is estimated exactly as TE
i
is estimated
in (5), with (V

i t
, U

i
) replacing (V
i t
, U
i
). Al-
though estimates of EE
i t
obtained in the rst
stage of the analysis donot take variation in ex-
5
A histogram of the temporal means of EE
i t
is provided in
gure 2, and the resulting similarity of OLS and tobit estimates
is veried in table 2.
6
Acomparison of MLEand OLS parameter estimates in table 2
illustrates the bias in the OLS parameter estimates.
planatory variables into consideration, the im-
pacts of these variables are incorporated into
estimates of CEE
t
. Since for any farm their
net impact can be either positive or negative,
CEE
i
> = < EE
i t
.
7
Determinants of Environmental
Efciency and Data
The main idea behind our approach to the
identication of explanatory variables is that
if we can use all the relevant relations in dairy
farming to compute the (rst-stage) efciency
scores, we will not nd any inefciency of ei-
ther type. Thus variation in efciency scores is
assumed to be caused by omitted variables and
measurement errors in the rst-stage analysis.
The estimated environmental efciency
frontier is compared with a comprehensive
model of dairy farming to identify the omit-
ted variables and measurement errors that are
expected to affect environmental efciency.
Figure 1 provides a scheme of the aforemen-
tioned model of dairy farming summarizing
all elements of dairy models from different
sciences and different levels of aggregation.
The model is described in detail in Reinhard.
Dairy farming consists of roughage produc-
tion and livestock production. The inputs
can be divided into variable inputs (in-
cluding nutrients), labor and capital. The
marketable output is mainly an aggregate of
milk and beef. The physical environment
consists of the exogenous physical factors re-
lated to land location (for instance weather,
soil quality). The institutional environment
inuences almost all inputs, outputs, and the
production process; for example, regulations
on the utilization of land, production quotas
for milk, levies on excess manure. Technol-
ogy is disaggregated into embodied and dis-
embodied technological change. The environ-
mental pressure captures the nutrient ows
from dairy farming into the natural environ-
ment. For example, nitrogen surplus consists,
among other things, of evaporation of ammo-
nia and leaching of nitrates into groundwater
(see Reinhard, Lovell and Thijssen).
We assume that calculated environmen-
tal inefciencies are due in part to omitted
7
EE
i t
in (7) varies across farms and through time, since it is a
function of X
i t
and N
i t
. In principle CEE
i
in (9) also varies across
farms and through time. However, the hypothesis that CEE
i
is
time-varying is rejected by the data, perhaps due to the temporally
short panel.
1058 November 2002 Amer. J. Agr. Econ.
Figure 1. A schematic of a comprehensive model of dairy farming
variables and measurement error. Thus the
factors depicted in gure 1 that are not mod-
eled appropriately in the rst stage must be
incorporated in the second stage. In the rst
stage we specify a stochastic production fron-
tier with a single output (an index of dairy farm
output), three conventional inputs (labor and
indexes of capital and variable inputs), a sin-
gle environmentally detrimental input (nitro-
gen surplus), and a time trend. Disembodied
technological change is captured by the time
trend, and embodied technological change is
captured by interactions of the time trend with
the conventional inputs. Environmental ef-
ciency is calculated, conditional on stochastic
disturbances, as in (7). This rst-stage analysis
is described in detail in Reinhard, Lovell and
Thijssen.
Two elements of the production process
are not taken into account in the rst stage
at all; namely the physical environment and
the institutional environment (see gure 1).
These elements are outside the control of the
farmer.
Another problem is the measurement of
the variables used in the rst stage. Variables
whose productive capacities are partly incor-
porated will cause apparent inefciencies. The
productive capacity of family labor in the rst
stage is expressed in hours. Labor quality is
not accounted for in this variable, and produc-
tive capacity depends on quality as well. La-
bor quality includes the farmers ability and
learning by doing. Variable inputs and out-
puts are aggregated from several components.
Information about quality is relatively well
Reinhard, Lovell, and Thijssen Analysis of Environmental Efficiency Variation 1059
preserved in this aggregation (Reinhard,
Lovell and Thijssen). However, information is
lost about the nitrogen content of the inputs
and the outputs. The quantity of nitrogen in
inputs and in outputs denes the nitrogen sur-
plus. To correct for this loss of information, we
use variables reecting nitrogen content of in-
put and output.
By using the aggregate capital stock vari-
able in the production frontier we implicitly
assume that the capital service ow of its
components is identical. Capital service ows
cannot be measured directly. Therefore, we
use additional information on capital services.
The physical environment captures differences
in physical environment (e.g., solar radiation,
soil, and infrastructure). The institutional en-
vironment is determined by government regu-
latory agencies and is outside the control of
the farmer. In the research period the im-
pact of Dutch environmental policy differs
across farms, because it depends on the phos-
phate surplus. In gure 1 factors that are not
modeled in the rst stage are marked only
with II, while factors that are not completely
modeled in the rst stage are marked with
I and II.
Technological development is assumed to
be captured in the rst stage. However, the
technology employed can differ across indi-
Table 1. The Explanatory Variables Used
Variable Mean Min Max SD
Labor quality
Agricultural education (4 categories)
Age of manager (years) 47.5 21 78 11.4
Years manager 20.2 1 54 11.1
Years FADN 2.7 1 9 1.9
Labor share of manager (%) 85.2 12.5 100 17.0
Off-farm income (NLG) 28,091 0 365,203 24,579
Nitrogen in inputs
Feed per cow (NLG/cow) 1110.8 122.6 6139.4 627.3
N fertilizer per hectare (kg/hectare) 256.1 0 569.9 84.7
Nitrogen in output
Dairy (%) 78.2 66.7 99.9 5.6
Capital specication
Number of cows 75.7 11 270 41.4
Sales and growth (NLG/cow) 743 641 2,850 258
Milk yield (kg/cow) 5315 2326 7673 868
Physical environment
Soil types (7-soil-type dummies)
Region (5 regional dummies)
Institutional environment
Year dummies (three-year dummies)
Note: The proportion of observations described by the dummy variables appears in table 2.
vidual farms. In a stochastic production fron-
tier setting, the economic behavior of farmers
is implicitly assumed to be either output or
prot maximization, conditional on the inputs.
The pursuit of alternative objectives or farm-
ing styles can also be a reason that farmers are
not on the frontier (Van der Ploeg, Renting
and Roex). Farming styles are not readily ob-
servable (Dijk et al.), and so they cannot be
used as explanatory variables. We might miss
minor relations as well, and also some mea-
surement error is still present in the variables
specied in the rst stage. Also the proxies we
use in the second stage do not capture these
factors completely. For instance regionandsoil
type are used to describe the physical envi-
ronment, but these two variables cannot cap-
ture all variation in solar radiation, precipita-
tion, etc. Therefore, we also have to deal with
an omitted variables problem in the second
stage. The variation of these omitted charac-
teristics in the second stage, U

, is captured
by the inefciency component of the second-
stage stochastic frontier model. The second-
stage regression model is expressed in log-
linear form, the variables appear in table 1,
and alternative estimation procedures are
used.
In table 1 we present the available explana-
tory variables in the Dutch Farm Accountancy
1060 November 2002 Amer. J. Agr. Econ.
Data Network (FADN).
8
We use off-farm in-
come as a proxy for the drive and motiva-
tion of the farmer. The abilities and capaci-
ties of the farmer are proxied by education,
which is distinguished in four categories. The
share of the managers labor in total family
labor is used to proxy the quality of family
labor (we assume that the manager is more
highly qualied than the other family mem-
bers). The background and experience is prox-
ied by the age of the manager, the number of
years of experience as a farm manager, and
the number of years of participation in FADN.
FADN participants receive an extensive bal-
ance sheet and nutrient account, and they have
reduced their nutrient surplus more than farm-
ers without a nutrient account (Poppe et al.,
p.78).
Nitrogen in inputs is proxied by the quan-
tity of feed bought per cow and the quantity
of nitrogen fertilizer per hectare. The quantity
of purchased feed per cow also incorporates
information about the presence of intensive
livestock. The share of dairy farming in total
production is an indicator of the specializa-
tion of the production process and is an indica-
tor of nitrogen in output. Differences between
intensive and extensive farming are included
in the variable feed bought per cow. Vari-
ation between specialized and mixed farm-
ing is captured by nitrogen in outputs. Capi-
tal specication is obtained by the size of the
herd and sales and growth per cow; the milk
yield is an indicator of the quality of the herd.
We do not use farm size as a separate vari-
able because it is strongly correlated with herd
size.
The physical environment is proxied by soil
type and regional dummies. The regional dum-
mies reect differences in solar radiation, wa-
ter availability, infrastructure, etc. Changes
in the institutional environment are proxied
by year dummies. We assume that regula-
tion affects all farmers identically. The year
dummies also incorporate annual differences
in weather conditions. All variables, Z
i t
, ex-
cept the dummy variables are normalized
by their sample means. The normalized vari-
ables are independent of units of measure-
ment, and the mean impact of each variable is
zero.
8
Abrief description of the Dutch FADN, and summary statistics
for therst-stagevariables (Y
i t
, X
i t
, N
i t
), canbefoundinReinhard,
Lovell and Thijssen.
An allocation rule whether to put the Z
j i t
in the rst stage or second-stage frontier does
not exist (to our knowledge). In the rst stage
production frontier inputs obeying the charac-
teristics of inputs (substitution, nondecreasing
relation with output) were used. In the second
stage, Z
j i t
specifying characteristics of the in-
puts (and output) are used that are expected
to inuence environmental efciency. No Z
j i t
is an input, according to the conventional def-
inition.
An Empirical Investigation into the
Determinants of Environmental Efciency
In this section we implement the second stage
of our two-stage model. In this stage, we quan-
tify the relationships between the explana-
toryvariables andtheenvironmental efciency
scores EE
i t
we obtained in the rst stage. Since
it is the environmental efciencies we wish
to explain, we summarize the distribution of
their temporal means in gure 2. With an over-
all sample mean of 0.441 and standard devia-
tion of 0.249, there is considerable variation in
environmental efciency to be explained in a
second-stage regression.
We used the software package FRONTIER
(Coelli) to generate maximum likelihood
estimates of the stochastic environmental ef-
ciency frontier specied in (8). We began with
the entire set of explanatory variables listed in
table 1. The contribution of each variable was
evaluated by computing the likelihood-ratio
test statistics. Variables not contributing
signicantly (at the 90% level) were deleted.
The dummy variables for soil and education
were aggregated into a smaller number of cat-
egories whenever the number of observations
in a category was small and the categories had
a comparable impact. We used one dummy
variable for education (1 =agricultural educa-
tion or higher education; 0 = otherwise), two
dummy variables for soil type (Soil 1 =1 if soil
type is sea sediment clay; 0 = otherwise; Soil
4 = 1 if soil type is sand; 0 = otherwise), and
one dummy variable for region (Region 3 = 1
if location is in eastern, middle and southern
livestock regions; 0 = otherwise). Number of
years of experience as farm manager, off-farm
income, growth and sales per cow and the
1992dummyvariable, werealsonot signicant,
and were deleted. The remaining explanatory
variables in the model and their parameter
Reinhard, Lovell, and Thijssen Analysis of Environmental Efficiency Variation 1061
Figure 2. Histogram of environmental efciency scores (613 farms)
estimates appear in the MLE column of
table 2.
9
We tested the appropriateness of the fron-
tier specicationby computing the skewness of
the OLS residuals. The

b
1
statistic (Schmidt
and Lin) is 0.76, indicating that the OLS
residuals do indeed exhibit the expected neg-
ative skewness, and that a stochastic frontier
model is appropriate. We tested the robust-
ness of the model by supplying starting values
differing from the OLS estimates, and found
the parameter estimates to be robust to alter-
native starting values. The half-normal restric-
tion on the truncated normal model was re-
jected, with a likelihood-ratio test statistic of
69.6 for the null hypothesis that = 0. The
estimated value of [=
2
U
/(
2
V
+
2
U
)] in-
dicates that environmental inefciency exists,
with a likelihood-ratio test statistic of 1,872 for
the null hypothesis that = 0. The relatively
large estimatedvalue of indicates that almost
the entire error inthe secondstage is due toun-
explained (by the explanatory variables) envi-
ronmental inefciency. The role of statistical
noise in explaining the original environmental
efciency scores is very small. The hypothesis
that the farm-specic conditional environmen-
9
Also a model was estimated in which two additional explana-
tory variables, feed per cow and nitrogen fertilizer per hectare,
were deleted. Although each is statistically signicant in the MLE
column, and a likelihood-ratio test rejects the hypothesis that they
can be deleted from the model, each is arguably not exogenous.
Fortunately, estimates and signicance levels of the remaining pa-
rameters are robust to the deletion of these two variables. The rank
correlation coefcient for the two sets of results is 0.981.
tal efciencies vary through time is rejected.
Thus we have 1,545 observations on EE
i t
but
only 613 observations on CEE
i
.
As a robustness check, the nal two columns
of table 2 report OLSandtobit estimates of the
parameters describing the environmental ef-
ciency frontier. As expected, OLS and tobit
parameter estimates are very similar, and both
are very different from the MLE parameter
estimates. These patterns provide compelling
evidence infavor of the use of a stochastic fron-
tier in the second stage.
The MLE parameter estimates presented in
table 2 convey two types of information: (i)
the impacts of the explanatory variables on
environmental efciency; and (ii) estimates of
conditional environmental efciency, obtained
from (9) as the ratio of observed to maximum
feasible environmental efciency.
The parameter estimates provide estimates
of partial elasticities of EE
i t
with respect to
each explanatory variable. For most variables
plausible expectations can be formed of the
signs of the partial elasticities, but we have
no expectations of their magnitudes. Agricul-
tural education or higher education (in con-
trast to no education or general education)
is positively (at 90%) related to environmen-
tal efciency. Experience, as measured by the
age of the farm manager, has a negative ef-
fect on EE
i t
. This is in line with Weersink,
Turvey and Godah (p. 453), who argue that
inexperienced farmers tend to be more knowl-
edgeable about recent (environment-friendly)
technological advances than are their older
1062 November 2002 Amer. J. Agr. Econ.
Table 2. Parameter Estimates for the Stochastic Environmental Efciency Frontier
MLE
OLS Tobit
Parameter Standard Parameter Parameter
Variables
a
estimate error estimate estimate
Constant 0.294

0.062 0.577 0.571


Labor quality
Agricultural education (0.89) 0.097 0.052 0.056 0.061
Age of the manager 0.099 0.057 0.468

0.464

Years in FADN 0.059

0.021 0.019 0.017


Manager share in family labor 0.109

0.045 0.070 0.070


Nitrogen in inputs
Feed per cow 0.209

0.035 0.476

0.472

N fertilizer per ha 0.162

0.035 0.171

0.171

Nitrogen in outputs
Percentage dairy 0.560

0.188 1.585

1.583

Capital specication
Number of cows 0.206

0.034 0.572

0.573

Milk yield 0.706

0.114 1.948

1.943

Physical environment
Soil 1 (0.13) 0.200

0.053 0.475

0.477

Soil 4 (0.49) 0.151

0.047 0.160

0.156

Region 3 (0.37) 0.171

0.049 0.211

0.220

Institutional environment
Dummy 1993 (0.26) 0.079

0.021 0.057 0.058


Dummy 1994 (0.25) 0.101

0.024 0.122

0.121

4.203

0.457
0.983

0.002

2
4.494

0.439
a
The parenthetical value behind the dummy variables indicates the percentage of the total observations that are described by each dummy
variable.
b
The parameter estimate differs signicantly from zero at the 95% level.
counterparts. Participation in the FADN has
a positive and signicant effect on EE
i t
, prob-
ably due to the knowledge farmers gain from
extensive balance sheets and nutrient accounts
provided by LEI. The ratio of the quantity of
labor by the manager(s) to total family labor
has a positive and signicant effect on EE
i t
.
The herd size has a negative and signicant ef-
fect on EE
i t
. Since more cows produce more
manure, this estimate seems logical. Remark-
ably, the herd size is positively (signicantly)
correlated to the technical efciency measure.
Although we selected highly specialized dairy
farms from the FADN, these farms can still
have other activities such as fattening hogs or
veal calves. Farms with a relatively high share
of dairy activities are signicantly more envi-
ronmentally efcient, because they are not in-
volved in activities producing a large nitrogen
surplus. Milkyieldis strongly positively related
to environmental efciency. Farms purchasing
considerable feed per cow (these farms are
also likely to have hogs or poultry) are signif-
icantly less environmentally efcient. Not sur-
prisingly, a large nitrogen fertilizer application
per hectare is signicantly and negatively re-
latedtoenvironmental efciency. The effect on
environmental efciency of having type 1 soil
(richinclay fromsea sediments) is signicantly
positive, since this is a fertile soil type. Con-
versely, type 4 soil (which is sandy) has a sig-
nicantly negative effect, since sand is the least
fertile soil type in the Netherlands. Location
in region 3 (a combination of eastern, middle,
and southern regions specialized in livestock
production) has a signicantly positive impact
on environmental efciency. A possible expla-
nation is that the concentration of livestock
farms inthese regions results ina relatively low
feed price. Finally, environmental efciency
exhibited signicant improvement in 1993 rel-
ative to 1991 and 1992, and again in 1994 rela-
tive to 1993 (in accordance with the results of
Reinhard, Lovell and Thijssen).
Explanatory variables with a large partial
elasticity (see table 2), and which can be
Reinhard, Lovell, and Thijssen Analysis of Environmental Efficiency Variation 1063
Table 3. Comparison of Environmental Efciency Scores EE
i
and
Conditional Environmental Efciency Scores CEE
i
(613 observations)
Mean Minimum 25th (%) Median 75th (%) Maximum
EE
i
0.434 0.00
+
0.239 0.429 0.621 0.95
CEE
i
0.572 0.00
+
0.374 0.626 0.793 0.95
inuenced relatively easily by the government,
are the main targets for designing policy. The
milk yield shows the largest elasticity. A more
productive breed of cows can be expected to
increase environmental efciency. This leads
also to a reduction of the herd, conditional
on the milk production, which will raise en-
vironmental efciency as well. The govern-
ment can improve the milk yield by encour-
aging genetic research. Reduction of feed per
cow and nitrogen fertilizer per hectare will in-
crease environmental efciency as well. While
our methodology does not provide guidance
on how to implement these reduced applica-
tions of variable inputs, technical extension
services can focus on suitable strategies to
reduce the amount of feed per cow and ni-
trogen fertilizer per hectare. The mineral ac-
counts that have been mandatory since 1998
for farms with more than 2.5 milk cows per
hectare (or equivalent livestock) will stimulate
farms to reduce their nitrogen consumption as
well, and provide insight into their nutrient ac-
counts. Young, well-educated farmers are the
more environmentally efcient farmers, andso
education of older farmers to acquaint them
with new environment-friendly technologies
is likely to increase environmental efciency.
Farmers participating for a longer period in
the FADN learn from the information they
receive in the FADN balance sheets and nu-
trients accounts. When they are provided with
extensive balance sheets and nutrient accounts
of their farms, farm managers can be expected
to learn how to improve their environmental
performance.
The unexplained shortfall of environmen-
tal efciency beneath best practice observed
in the sample is reected in U

. Conditional
environmental efciency scores CEE
i
, ad-
justed for the explanatory variables charac-
terizing each farm, are computed from U

i
using (9). They provide estimates of environ-
mental efciency, conditional on the explana-
tory variables. Whereas in the rst stage a farm
may be penalized for its unfavorable circum-
stances, these factors are accounted for in the
second stage. Farm managers may rightly ob-
ject to their EE
i t
scores, which do not incorpo-
rate exogenous inuences, but they have less
reason to object to their CEE
i
scores, which
do incorporate exogenous inuences.
First-stage temporal means of EE
i t
scores
(noted as EE
i
) and second-stage CEE
i
scores
are summarized in table 3. The sample mean
of the CEE
i
is 0.572 with a minimum value of
0.00
+
and a maximum value of 0.95. In 94%
of the observations CEE
i
> EE
i
. The rank
correlation of the rst-stage EE
i
scores and
the second-stage CEE
i
scores is 0.926. The
dispersion of the EE
i
scores and the CEE
i
scores is of similar magnitude, with sample
standard deviations of 0.235 and 0.252, respec-
tively. The distribution of the conditional envi-
ronmental efciency scores appears in gure 3.
The distribution differs in large part from the
distribution of the rst-stage environmental
efciency scores because no distribution is
imposed on the EE
i t
scores while the CEE
i
scores are assumed to follow a truncated nor-
mal distribution.
The adjustment to EE
i t
scores by the ex-
planatory variables leads to generally higher
CEE
i
scores. The unexplained part of EE
i t
re-
ected by CEE
i
is due to a number of factors.
We had to use proxies to model the factors
omittedinthe rst stage or not accurately mea-
sured. We could not incorporate all relevant
information in the second stage (e.g., farming
styles). We modeled only the most important
factors; for instance solar radiation, temper-
ature, and water are not modeled explicitly.
The rst-stage environmental efciency scores
may not accurately reect environmental ef-
ciency (Reinhard, Lovell, and Thijssen, p. 56,
footnote 11).
The magnitude of the CEE
i
gives an indica-
tionof theproblemconfrontingenvironmental
policy. We do not have explanatory variables
available to clarify this portion of environ-
mental efciency. Therefore, we cannot read-
ily provide instruments for policy makers to
reduce CEE
i
. Complementary methods (e.g.,
visiting farms) may provide more insight into
the factors determining CEE
i
. For instance,
farmers can be interviewed to obtain more
1064 November 2002 Amer. J. Agr. Econ.
Figure 3. Histogram of conditional environmental efciency scores (613 farms)
information about their objectives and moti-
vation, and other elements of dairy farming
that cannot be captured in a balance sheet.
Summary and Conclusions
We have developed an analytical framework
within which to estimate the impact of var-
ious explanatory variables on environmental
efciency as a second-stage stochastic environ-
mental efciency frontier. The environmen-
tal efciency scores were computed in a rst-
stage analysis. Our second stage differs from
other approaches found in the literature, be-
cause we apply a stochastic frontier in the sec-
ond stage. The second-stage parameter esti-
mates reect impacts of explanatory variables
that can guide policy to increase environmen-
tal efciency. This methodology also generates
estimates of conditional environmental ef-
ciency that identify farms with relatively high
and relatively low environmental efciency,
conditional on their explanatory variables. We
showed that the second stage can be imple-
mented empirically, by estimating conditional
environmental efciency scores for each farm
in a panel of 613 Dutch dairy farms during the
199194 period.
The mean conditional environmental ef-
ciency is higher than the rst-stage environ-
mental efciency because we explain a por-
tion of the environmental efciency with the
explanatory variables (among others, indica-
tors of labor quality and the physical environ-
ment). We found that insight into the nutrient
balance, and the milk yield each affects the en-
vironmental efciency score positively.
[Received January 2000; nal revision received
September 2001.]
References
Battese, G.E., and T.J. Coelli. A Model for Techni-
cal Inefciency Effects in a Stochastic Frontier
Production Function for Panel Data. Empir.
Econ. 20(1995):32532.
. Prediction of Firm-Level Technical Ef-
ciencies with a Generalized Frontier Produc-
tionFunctionandPanel Data. J. Econometrics
38(July 1988):38799.
Coelli, T.J. A Guide to FRONTIER, Version 4.1:
A Computer Program for Stochastic Frontier
Production and Cost Function Estimation.
University of New England, Armidale, NSW,
Australia, 1994.
Dijk, J., C. Baarda, R. van Broekhuizen, T. de Haan,
W. Hennen, J.D. van der Ploeg and G. van de
Ven, Ins and Outs; Amultidisciplinary study af-
ter input-output relations and their relation with
the decision making of farmers (in Dutch). The
Hague: Agricultural Economics Research In-
stitute (LEI), Onderzoekverslag 160, 1998.
Fox, G., P.A. Bergen and E. Dickson. Why are
Some Farms More Successful than Others? A
Review. Size, Structure, andthe Changing Face
of American Agriculture. A. Hallam, ed. Boul-
der CO: Westview Press, 1993.
Reinhard, Lovell, and Thijssen Analysis of Environmental Efficiency Variation 1065
Hallam, A., and F. Machado. Efciency Analy-
sis with Panel Data: A Study of Portuguese
Dairy Farms. Eur. Rev. Agr. Econ. 23:1(1996):
7993.
Kumbhakar, S.C., and C.A.K. Lovell. Stochas-
tic Frontier Analysis. New York: Cambridge
University Press, 2000.
Pitt, M.M., and L.-F. Lee. The Measurement and
Sources of Technical Inefciency in the In-
donesian Weaving Industry. J. Develop. Econ.
9(August 1981):4364.
Poppe, K.J., F.M. Brouwer, J.P.P.J. Welten and
J.H.M. Wijnands eds. Agriculture, Environment
and Economics (in Dutch). The Hague: Agri-
cultural Economics Research Institute (LEI),
PR 6892, 1995.
Reinhard, S. Econometric Analysis of Eco-
nomic and Environmental Efciency of Dutch
Dairy Farms. PhD dissertation, Wageningen
University, 1999.
Reinhard, S., C.A.K. Lovell and G. Thijssen.
Econometric Estimation of Technical and
Environmental Efciency: An Application to
Dutch Dairy Farms. Amer. J. Agr. Econ.
81(February 1999):4460.
Reinhard, S., and G. Thijssen. Nitrogen Efciency
of Dutch Dairy Farms; A Shadow Cost Sys-
temApproach. Eur. Rev. Agr. Econ. 27(2000):
16786.
Schmidt, P., and T-F. Lin. Simple Tests of Alter-
native Specications in Stochastic Frontier
Models. J. Econometrics 24(March 1984):349
61.
Sitorus, B.L. Productive Efciency and Redundant
Factors of Production in Traditional Agri-
culture of Underdeveloped Countries: A
Note on Measurement. Proceedings of the
thirty ninth annual meeting of the West-
ern Farm Economics Association, pp. 15358,
1966.
Timmer, C.P. Using a Probabilistic Frontier
Production Function to Measure Technical
Efciency. J. Polit. Econ. 79(JulyAugust
1971):77694.
Van der Ploeg, J.D., H. Renting and J. Roex. Mul-
tiple Comparisons and a lot of Unknowns:
an Exploring Research After Empirical Input-
Output Relations in Dutch Dairy Farming (in
Dutch). The Hague: NRLO, Rapport 94/1,
1994.
Weersink, A., C.G. Turvey and A. Godah.
Decomposition Measures of Technical Ef-
ciency for Ontario Dairy Farms. Can. J. Agr.
Econ. 38(November 1990):43956.

You might also like