Академический Документы
Профессиональный Документы
Культура Документы
3, 2014, 579599
doi: 10.1111/1477-9552.12057
Abstract
This paper examines the relationships between the marginal abatement costs
(MAC) of greenhouse gas (GHG) emissions on dairy farms and factors such as
herd size, milk yield and available farm labour, on the one hand, and prices, GHG
indicators and GHG reduction levels, on the other. A two-stage Heckman procedure is used to estimate these relationships from a systematically designed set of
simulations with a highly detailed mixed integer bio-economic farm-level model.
The resulting meta-models are then used to analyse how MAC vary across farmlevel conditions and GHG measures. We nd that simpler GHG indicators lead to
signicantly higher MAC, and that MAC strongly increase beyond a 15% emission reduction, depending on farm attributes and the chosen indicator. MAC
decrease rapidly with increasing farm size, but the eect levels o beyond a herd
size of 40 cows. As expected, the main factors driving gross margins per dairy cow
also signicantly inuence mitigation costs. Our results indicate high variability of
MAC on real life farms. In contrast to time consuming simulations with the complex mixed integer bio-economic programming model, the meta-models allow the
distribution of MAC in a farm population to be eciently derived and thus could
be used to upscale to regional or sector level.
Keywords: Dairy farms; greenhouse gas indicators; Latin-Hypercube sampling;
marginal abatement costs; meta-modelling.
JEL classifications: Q12, Q52, C34, C99.
Bernd Lengers is a Research Assistant at the Institute of Food and Resource Economics, University of Bonn, Germany. E-mail: bernd.lengers@ilr.uni-bonn.de for correspondence. Wolfgang Britz is a Researcher and Karin Holm-M
uller is Professor at the same Institute. The
research is funded by a grant of the German Science Foundation (DFG) with the reference
number HO 3780/2-1. The authors thank anonymous referees for their comments on an earlier
version of this paper.
580
1. Introduction
The mitigation of greenhouse gases (GHGs) from agricultural production processes is
broadly discussed from a political and a scientic viewpoint, both with regard to
abatement costs (AC) (Anvec, 2011; DeCara and Vermont, 2011) and whether and
how to incorporate agriculture into GHG reduction eorts (e.g. Ramilan et al.,
2011). Any decision on policy instruments and potential reduction levels requires
knowledge about possible abatement options, abatement potentials and, in particular,
related abatement costs. For political purposes, knowledge about the marginal abatement costs (MAC) is of high relevance as these costs determine the mitigation potential under each price-based GHG mitigation policy. Farmers will only mitigate GHGs
as long as MAC for GHGs are below the CO2-equivalent prices.
Various studies (see section 2) are available regarding possible abatement strategies
and potential GHG savings from agriculture, oering a wider range of MAC estimates. However, these are hard to generalise as they depend on factors not systematically controlled for, such as exogenous assumptions on prices, farm characteristics,
the model or estimation structure or the GHG indicator used. A GHG indicator is
necessary as an accounting scheme for GHG quantication, since direct measurements are not possible on farms. Published meta-analyses on MAC by Vermont and
DeCara (2010) as well as Kuik et al. (2009) only investigate dierences in methodological approaches and assumptions. To our knowledge, no study exists which systematically analyses drivers of dierences in MAC and thus farm income changes
provoked by GHG-related policy instruments. Yet this is certainly a signicant aspect
of the policy debate about inclusion of agriculture in emission reduction eorts. We
develop and illustrate a methodology for meta-analyses and apply it to dairy farm
conditions in a larger German region to provide an analysis of these issues.
The evidence, for example from Lengers and Britz (2012), about dierences in
MAC between farms and the necessity of GHG reduction policies to use GHG indicators (Scheele et al., 1993, p. 298), gives rise to two major questions. (1) Which are the
most important farm characteristics impacting MAC and what are their quantitative
eects? and (2) What is the relationship between the applied GHG indicator and other
drivers such as prices and MAC?
Answering these questions requires a set of farm-level observations of MAC with sucient variation in key factors, but time and cost considerations exclude real-life experiments for a reasonable group of existing farms. But even simulating abatement strategies
with a computer model for a larger set of farms can be a time consuming exercise. We
therefore conduct a representative sample of what-if simulations following DOEprinciples (design of experiments: Kleijnen, 1999, 2005) with the single farm model
DAIRYDYN, a highly detailed bio-economic model for specialised dairy farms. The focus
on dairy farms is motivated by their high share in German agricultural GHG emissions.
From these simulation results we derive meta-models for a simple and a rather
detailed GHG indicator which allow us to analyse the relationships between MAC
and key factors such as predened characteristics of dairy farms, prices and GHG
reduction level. The meta-models summarise the behaviour of the underlying, more
complex simulation model and show how MAC depend on these factors and the
applied GHG indicator over the range of varying circumstances of real farms. Our
ndings complement the existing literature which typically provides results either for
selected single farms only, or at a rather high aggregation level of larger administrative regions, but cannot give farm-specic information on what drives MAC.
581
The organisation of the paper is as follows: Section 2 briey discusses available literature, including reported ranges of MAC. In the Material and Methods section, we rst
give an overview about our methodology, before we briey outline DAIRYDYN and
present the dierent GHG indicators for which MAC will be investigated. Next, we discuss the set up of our experiments single farm simulation runs which cover relevant
dairy production systems in North Rhine-Westphalia, Germany, based on an ecient
space lling sampling procedure recognising conditioning factor correlations. Subsequently, we discuss the estimation of the meta-models for two dierent GHG indicators
analysed based on a Heckman two-stage selection procedure. The statistical estimator
delivers both test statistics and error terms for the given sample and shows its validity
regarding the sampling background. Simulations with the meta-models are used to
present the main ndings in the results section before we summarise and conclude.
2. Literature Overview
Various studies are available regarding possible abatement strategies and potential
GHG savings from agriculture (e.g. Olesen et al., 2004, 2006; Flachowsky and Brade,
2007; Smith et al., 2008; Niggli et al., 2009), and related abatement and marginal
abatement costs. Whereas MAC are generally dened as the costs required to abate
an additional unit of GHG, the approaches to quantify them dier considerably. The
studies concerning MAC can be broadly categorised in three groups (Vermont and
DeCara, 2010). The rst group uses so-called engineering models (e.g. Weiske and
Michel, 2007; Moran et al., 2011) which rank existing abatement options according to
their abatement potential by increasing per unit abatement costs. Interactions between
the options are typically not recognised.
The second group uses aggregate programming models (e.g. CAPRI by Perez,
2006; AROPAj by DeCara and Jayet, 2000) for farm type groups or regions covering
the EU, often with less engineering detail. The sole or combined use of equilibrium
models, such as with CAPRI by Perez (2006) or with a global CGE framework in
Golub et al. (2009), allows consideration of price feedback and further market interactions. Equilibrium model approaches tend to nd higher MACs compared to
engineering models for two reasons. Firstly, induced output price increases raise
MAC further, and secondly, these models typically do not consider all low-cost abatement options analysed in engineering models.
The third group, which comes close to our approach, are supply side models with a
rich technology description (e.g. DeCara et al., 2005; Durandeau et al., 2010; Ramilan et al., 2011; Lengers and Britz, 2012). These studies in most cases only concern
specic regions or single farm types. While typically considering the abatement
options also covered by engineering approaches, they optimise the farms plan under
an emission ceiling or a given CO2 permit price. This allows the simultaneous introduction of several abatement options while recognising their potential interactions.
Studies with supply-side models tend to nd the lowest MAC of the three groups.
Compared to engineering models on the one hand, the prot maximal ranking and
combination of abatement options lowers their MAC estimates. On the other hand,
supply-side models typically also consider low-cost abatement options not covered by
more aggregate approaches. However, the lower MAC found by supply-side models
are also the outcome of neglecting induced output price increases.
In a recent review of agricultural MAC derived by dierent approaches, Vermont
and DeCara (2010) found a high variability in estimated MAC between 0.19 and
2014 The Agricultural Economics Society
582
535.80 ton1 CO2-equ. for abatement rates of 068% of baseline emissions, and also
show large dierences in MAC for the same percentage GHG reductions (Vermont
and DeCara, 2010, pp. 13831384). They conclude that the observed variability in
MAC estimates is rooted to a large extent in what type of approach is used, as stated
above. More generally, Barker et al. (2002) and Kuik et al. (2009) show that MAC
derived by simulation models depend on their structural characteristics and further
assumptions such as the emission baseline used and the relevant time interval for
emission quantication. Lengers et al. (2013a) stress again the inuence of the chosen
abatement level, and, in addition, together with Lengers and Britz (2012) highlight the
importance of the GHG indicator for the estimated MAC, a point often neglected in
other studies. They argue that the chosen abatement strategies and consequently
MAC strongly depend on the GHG indicator as farmers can only be expected to
adopt abatement options which are credited. Besides that, an economic perspective
suggests that MAC should clearly depend on prices of input and output and further
characteristics of the farms investigated, points often neglected in existing studies.
3. Material and Methods
3.1. Overview on methodology
The literature review suggests that single farm approaches are best suited to reect differences related to farm attributes and GHG indicator choice as they depict the complex bio-physical and bio-economic processes in agricultural production in sucient
detail. However, ndings for a specic single farm are dicult to generalise to more
general farm types or regions (Stoker, 1993), the level of interest for policy decisions.
Conducting simulations to cover the distribution of relevant factors in the farm population can consume time and resources. For example, simulating MAC for just 70
dairy farms over a 15-year planning horizon with the DAIRYDYN model used in
Lengers et al. (2013a) took more than 3 days with an 8-core processor. Runtime considerations are thus important as they restrict the number of possible experiments
(Bouzaher et al., 1993, p. 3; Carriquiry et al., 1998). Meta-modelling seems thus inviting as it can replace time consuming simulations using a complex computer model
with a far simpler one, helping to overcome computational restrictions. A meta-model
approximates the output (response) of the more complex model using standard statistical techniques on model results from representative variations of factors in the
underlying complex model. It identies the most important factors for model results
and leads to a simpler functional form with fewer input variables (factors) (Carriquiry
et al., 1998, p. 507). A meta-model thus quanties major inputoutput (I/O) relationships embedded in the structure of the more complex model (Kleijnen, 2008) and can
therefore improve our understanding of real-life systems (Bouzaher et al., 1993, p. 3).
Consequently, the rst step in the development of a meta-model is the generation of
a set of model results. In order to account for the interactions of dierent factors2
such as farm attributes, prices and indicator choice on MAC, we construct a larger set
of simulations where factor levels are systematically and simultaneously3 varied, also
2
583
respecting possible correlations between the variables. Simulation results are then
used to derive appropriate statistical meta-models. The systematic of the overall
approach is visualised in the following gure.
DAIRYDYN
Meta-Modelling
In the following, we discuss the dierent elements shown in Figure 1. The set-up of
the experiments is clearly related to the structure of the simulation model such that we
rst briey present the DAIRYDYN model (section 3.2). Next, we discuss which
factors are varied (section 3.3) and how the experiments are designed (section 3.4).
Finally, we discuss the statistical estimator.
3.2. The simulation model DAIRYDYN
DAIRYDYN (Lengers and Britz, 2012) is a highly detailed, fully dynamic mixed integer linear programming model (MIP) for the simulation of economically optimal
farm-level plans on specialised dairy farms, realised in GAMS (general algebraic modelling system). It maximises the expected net-present value over several years and different states of nature under GHG emission ceilings related to specic GHG
indicators. The model hence implements full technical and allocative eciency, which
means that so-called winwin abatement options which reduce emissions and
simultaneously raise farm prots are already realised in the unrestricted baseline.
It takes into account investment and labour use decisions, respecting their noncontinuous character, and considers sunk costs of past investment decisions along
with the evolving path dependencies. Specically, existing stables cannot be sold or
rented out, and the resale value of machinery is assumed to be low. DAIRYDYN is
based on a detailed production-based approach, simulating farm management
decisions and related material ows for animal husbandry, cultivation of land, feed
production and feeding as well as manure management, partly on a monthly basis.
2014 The Agricultural Economics Society
584
The simulation model is designed to analyse impacts of GHG emission ceilings compared with a baseline of an optimised farm plan without any emission ceiling. The
fully dynamic, forward-looking character of DAIRYDYN excludes a formal calibration against existing observations, since the time series with the necessary detail are
missing. Instead, we rely on the combination of highly detailed information from farm
management handbooks (KTBL, 2010) and a rich set of constraints. However, the
GHG emission inventory simulated with the model for a specic farm can be compared against long-term measurements (Lengers et al., 2013b).
The decision variables in the simulation model are linked to a GHG accounting module
which incorporates ve dierent Intergovernmental Panel on Climate Change (IPCC)based GHG calculation schemes (indicators) to quantify the farm-level GHG inventory
(IPCC, 2006; Lengers, 2012). These GHG indicators dier in the level of detail of required
process information. In our analysis, we only use the two extreme indicators, i.e. the simplest (actBased) based on default per animal or per ha emission parameters (comparable
to the IPCC Tier 1 approach) and the most detailed (refInd), which implements highly
detailed process information in its GHG calculation such as, for example, milk yield level,
composition of the ration, type and duration of manure storage, synthetic and organic fertiliser practice and manure coverage techniques, at a yearly and, where necessary, monthly
resolution (derived from the IPCC Tier 3 approach) (see Lengers and Britz, 2012; Lengers
et al., 2013a for more detail, or Lengers, 2012 for a full description). We omit the remaining three intermediate GHG indicators as earlier results showed that abatement strategies
and related costs for these intermediate indicators do not dier signicantly from one or
other of the two extremes (Lengers and Britz, 2012; Lengers et al., 2013a). Whereas the
simple indicator (actBased) is only sensitive to changes in activity levels (change of animal
numbers, cropping and grassland acreage decisions), the advanced indicator (refInd) also
accounts for changes in the intensity of management, feed composition, fertiliser practice
and manure storage types and time (Lengers, 2012). This leads to dierent abatement
strategies and associated costs under the GHG indicators, as shown by Lengers and Britz
(2012) because dierent GHG mitigation options (permanent and variable ones) are
accounted by the dierent detailed indicators (see Table 1).
Table 1
GHG abatement options recognised by indicators
actBased
Permanent
Manure management techniques
Application techniques
Variable
Fodder optimization
Breeding activities
Intensity management
N-reduced feeding
Fertiliser practice
Area cultivated
Herd size managment, crop growing decisions
Feed additives/fat content
Pasture management/ increase grazing
X
X
X
X
refInd
X*
X
X
X
X
X
X
X
X
585
586
587
Table 2
Overview of explanatory factors
Variable/attribute
Name
Unit
Min
Median
135
Number of cows*
nCows
Head
20
Milk yield
milkYield
kg ECM
cow1 a1
5,000 8,000
Age of stables
StableAge
Labour
productivity
Building year
1995
of the stable
WHperCow Working hours 30
cow1 a1
Milk price
milkPrice
Wage rate
Max
Data source
250
2000
IT.NRW, 2012;
LKV-NRW,
2012;
11,000 LWK-NRW,
20082012;
LKV-NRW,
2012;
2005
51
72
25.97 31.205
36.44
wageRate
-cent kg
ECM1
hour1
10.5
15
Price concentrate
concPrice
ton1
160
195
230
Time horizon
GHG reduction
level
horizon
Years
%
10
1
15
10.5
20
20
LWK-NRW,
20082012;
FDZ, 2013;
BMELV,
19912011, 2012
LFL, 2012;
LWK-NRW,
2012
adapted to
KTBL, 2010;
LFL, 2012;
LWK-NRW,
20082012
combinations with DAIRYDYN.6 Therefore, a limited but representative set of factor level combinations is selected based on DOE principles (Kleijnen, 1999). Specically, we apply a Latin-hypercube sampling (LHS) method, which is more ecient
than simple random sampling (McKay et al., 1979; Iman and Conover, 1980; Iman
et al., 1981, pp. 176177; Giunta et al., 2003, p. 7; Iman, 2008). LHS denes a number
of experiments which simultaneously change levels of various factors, while being representative for the full range of possible factor level permutations. Based on space lling designs, LHS smoothly samples over the k-dimensional input space for a dened
size of the sample (Iman et al., 1981, p. 176; Owen, 1992, pp. 443445). It does not
necessitate a decision beforehand about for which factors a more ne-grained
6
All GHG reduction levels are simulated in each experiment, which leaves us with eight factors.
If we would allow for ve dierent levels for each factor and simulate all combinations, we
would need to conduct 5^8~400.000 experiments.
588
resolution of levels is appropriate (Iman, 2008). However, standard LHS assumes zero
correlation between the factors which may lead to invalid statistics compiled from the
output if factors are correlated in reality (Iman and Conover, 1982, p. 331). We thus
employ a LHS procedure according to Iman and Conover (1982) which considers factor correlations.
We used the known correlations and factor value ranges (see section 3.3) to determine a representative sample. To do so, we employed the statistical software R (version 2.15.1) for the DOE generation, specically the LHS-package lhs_0.10 in
combination with the algorithm from Iman and Conover (1982) to incorporate correlations. As the accuracy to which that algorithm can recover the correlation depends
on the drawn sample, we performed the LHS a few thousand times for a sample size n
of 200 and selected the LHS-sample with the best t between the randomised and the
given correlation matrix (LHS runs stopped when the dierence between drawn and
assumed correlation matrix did not decrease further with additional LHS runs). That
sample size was chosen to allow running the simulations in an acceptable time, while
still generating a suciently large sample for the econometric estimation of the MAC
for the two GHG indicators.
Figure 2. Normalised scatter plot matrix of 200 LHS draws for eight factors
589
The histograms of the sampling outcome in the diagonal boxes in Figure 2 suggest
the desired uniform distribution over the normalised [01] factor range (cf. Wyss and
Jorgensen, 1998, p. 7) for each factor. The normalised scatter plots below the diagonal
depict both the space lling design as well as showing the drawn correlation between
factor pairs. The matching boxes across the diagonal show a satisfactory t between
the drawn (without parentheses) and the given (in parentheses) correlation factors.
The improved LHS is hence ecient as it is space lling and shows a good t to the
given factor correlations (the mean percentage deviation of given and drawn correlations of the whole random sample is 7.76%), despite a relatively low number of draws
as required by Carriquiry et al. (1998, p. 507).
3.5. Model runs
We excluded some draws with implausible factor level combinations (e.g. very high
milk yields in combination with very small herd sizes), leaving us with 155 single farm
experiments. Each experiment comprised 20 dierent optimisation runs, with increasing GHG reductions from 1% to 20% of the baseline emissions. The exact reduction
level was randomised between +/ 0.5 percentage points around these reduction steps.
Each experiment was conducted for the two GHG indicators, resulting in 3,100 observations on MAC for each indicator.
3.6. Statistical meta-modelling
In order to quantify which factors signicantly impact MAC under the two GHG
indicators and to visualise the simulations, a statistical meta-model is constructed for
each of them. These statistical response surfaces can be understood as approximation
of the simulation programmes I/O transformation (Kleijnen, 1999, p. 116) in this
study for MAC depending on farm attributes, prices, GHG reduction levels, time
horizon and the GHG indicator applied.
This requires selecting an appropriate statistical estimator. As our non-standard
LHS which accounts for factor correlations is no longer orthogonal, we therefore rst
tested for multicollinearity of the randomised sample variables (not critical, the variance ination factors were below 10).7
Second, a graphical analysis of results revealed that in certain experiments, farms
exit dairy production at more aggressive GHG emission ceilings. The exit decision is
mainly driven by the opportunity cost of labour (returns to labour at optimal farm
programme compared with o-farm wage) and land (a xed, but rather low land (rental) price) which stays unchanged in all simulation runs), which depend in turn on
farm characteristics and prices for outputs and inputs. Once a farm has exited, its
GHG emissions will stay unchanged at zero and prots will no longer change. Hence,
further tightening of the emission ceiling cannot deliver useful information on MAC.
Leaving these zero observations in the sample would lead to biased results, while
excluding them would omit information (censored data) and cause a sample selection
bias8 (Heckman, 1979; Kennedy, 2008, p. 265). Therefore, we apply a two-stage
Marquardt (1970) suggests that serious collinearity is present for VIF-values above 10.
Tobin (1958) rst showed that if censoring of the dependent variable is not considered in the
regression analysis, an ordinary least squares (OLS) estimation will produce biased estimates.
8
590
Heckman estimation procedure which in the selection equation rst estimates the
probability of exiting farming (probit-model) and in the outcome equation (OLS linear regression), conditioned on the rst stage probability, the MAC. Accordingly, as
standard in the Heckman approach, the inverse Mills ratio calculated from the rst
stage probit-model was added as an explanatory variable to the second stage to correct for the self selection bias.9 (Heckman, 1979; Kennedy, 2008, pp. 265267) The
estimation is performed with the R SampleSelection package (Toomet and Henningsen, 2008).
There is a second type of observation with zero MAC. Linear programming models
might react with a basis change and thus non-smooth reactions to changes in binding
constraints. This behaviour is reinforced, as in DAIRYDYN, by the presence of integer variables. Due to a basis change, introducing a GHG emission ceiling might lead
to a higher reduction than required. In our stepwise reduction simulations, to give an
example, a 1% increase in the enforcement level could lead to a reduction in GHGs of
more than 2%. In this case, if the next reduction is smaller than 1%, it will not require
adjustments of the farm programme. Thus, prots and abatement costs might not
change in a consecutive step for which, accordingly, the MAC become zero. These
zero MAC observations are kept in the sample, but clearly will reduce the explained
variance of the outcome equation, which, as a multiple linear regression model, will
react smoothly to changes in explanatory variables such as the GHG emission ceiling
and cannot generate the kind of jumpy MAC curves which certain experiments might
simulate.
There is no reason to assume a priori that relationships between MAC and farm
attributes, as well as further explanatory variables such as prices should be linear, or
to exclude interaction eects in the design matrix. Hence, after some tests, we rst
added the square root and the square of the reciprocal value of each factor to the
design matrix. Similarly, we introduced interaction terms between all untransformed
factors as explanatory variables. The resulting design matrix tends to be highly co-linear. Therefore, in a next step we estimated a linear regression between each explanatory variable and all others, dropping any explanatory variable with a multiple
correlation above 94%. Afterwards, we used a backward selection strategy, dropping
any variable with a signicance level above 5% from the Heckman model.
4. Results
The selected explanatory variables with their estimated coecients for the two metamodels are shown in the Appendix (see the additional supporting information accompanying the online version of this article). The non-linear transformations of the
factors and the presence of interaction terms render a direct interpretation of the estimated coecients challenging.10 In order to visualise the eect of single factors on the
9
Inverse Mills ratio is added in order not to omit information of the explanatory variables of
cases censored in the selection step. An explanation of this step is also given in Giovanopoulou
et al. (2011, p. 2177).
10
Imagine
the
following
functional
form
of
a
meta
model:
MAC
2
b0 b1 x1 b2 1=x1 b3 x1 x2 . In order to assess the impact of factor x1 on the estimated
MAC, one needs to consider simultaneously all terms comprising x1, i.e. @MAC =@x1 which is
not directly possible from examining the estimated coecients, as it requires consideration of
not only the non-linearities, but also the level of other factors, in the example of x2.
591
MAC, we therefore separately varied the value of each factor over its range (cf.
Table 2), while xing all remaining factors at their median values. From there, we
generated a set of new observations based on the transformations and interaction
terms present in the estimated output equations (see Appendix S1 and S2, found in
the additional supporting information accompanying the online version of this article)
for which we simulated the meta-models. This allows a graphical depiction of the
dependence of MAC on changes of each factor.
The resulting graphs (Figure 3) indicate that MAC dier considerably between
GHG indicators. Also, how they are aected by the dierent factors is crucially dependent on the GHG indicator chosen. The gures show that the simple indicator provokes higher MAC in comparison with the more advanced indicator, which accounts
for more exible and low-cost abatement strategies (farm size reductions are the only
options accounted for by the simple indicator as described in section 3).
For all graphs, a steeper curve means a greater eect of factor level changes on the
MAC over the factor range considered in the experiments. Hence, for both GHG indicators the GHG reduction level (between 1% and 20%) has the highest impact. As for
the estimates under the advanced indicator (rst graph of Figure 3), MAC increase
with diminishing rates if the abatement level increases, the simple indicator induces
MAC curves which stay constant for intermediate reduction levels as the farm basically shrinks proportionally (reduction of cow numbers and acreages with constant
losses of gross margins). For larger reductions above 15% under the simple indicator,
the MAC increase as it becomes more likely in the sample to work half or full-time
o-farm, which leads to an increase in opportunity costs of labour as the eective
wage rate increases for half or full-time work. That nding shows the importance of
reecting indivisibilities of labour in the integer approach underlying the simulation
model.
Dierences between the GHG indicators are mostly found for low reduction levels.
For a 1% reduction the meta-model of the advanced indicator estimates MAC near
0 whereas the meta-model of the simple indicator estimates MAC about 50 ton1
CO2-equ. The advanced indicator also allows for low-cost abatement options as long
as the abatement potential of low-cost measures is not fully utilised. Here, during the
rst 15% reduction of baseline emissions, the highest increase in MAC takes place,
where MAC easily double between reduction steps. This shows that cheap abatement
strategies are used rst, with however only a limited reduction potential. At these
moderate reduction levels, the advanced GHG indicator benets from oering lowcost abatement options not credited under the simple indicator; accordingly, its MAC
are considerably below those induced by the simple indicator. For higher reduction
levels MAC under both indicators reach a comparable level where abatement strategies under both GHG indicators consist of farm size reductions. This triggers the
observed steep increase in MAC once the abatement potential of low-cost measures is
fully utilised and explains the decline in abatement cost advantages under the
advanced indicator with increasing abatement level.
The eect of the milk price shows that MAC are quite output price sensitive, independent of the indicator used. This can be easily understood from the fact that over a
larger range of the GHG emission reductions, abatement eorts are linked to output
reductions.
The eect induced by the herd size (number of cows in initial herd) of the farms is
quite high for smaller farms below 40 cows. For larger farms, there seems to be no signicant variation in MAC due to herd size variations. We interpret this nding by the
2014 The Agricultural Economics Society
592
250
200
150
100
simulated by meta-model of simple indicator
50
0
1
11
15
20
250
200
150
100
50
0
25.97
250
200
150
100
50
0
28.59
31.21
33.82
20
36.44
78
250
200
150
100
50
0
5,000
135
200
150
100
50
0
6,500
8,000
9,500
11,000
10
11
12
13
14
15
16
17
18
19
20
250
200
150
100
50
0
1995
250
200
150
100
50
0
2000
2005
6.0
8.3
10.5
12.8
15.0
213
230
250
250
200
150
100
50
0
30.0
193
250
200
150
100
50
0
40.5
51.0
61.5
72.0
160
178
195
Figure 3. MAC in ton1 CO2-equ. simulated with meta-models depending on factor levels
Note: All other factors are xed at their median; Mills-ratio from probit included.
593
fact that investment-based mitigation options such as manure coverage, which allow
for low-cost abatement, are not realised on small-scale farms due to economies of
scale, which however quickly level o.
Higher milk yields boost economic returns per cow and also per GHG emitted and
thus provoke MAC increases under both indicators. The eect is less pronounced for
the advanced indicator (curve is less steep) for two reasons: rst, the simple indicator
assigns a default emission factor to each cow, independent of the output level, whereas
the advanced indicator also recognises the diminishing GHG emissions per kg of milk
with increasing output level per cow. It hence requires lower herd size reductions
compared with the simple indicator. Second, as the advanced indicator also accounts
for feed ration changes which lower emissions, stricter emission ceilings (hence higher
GHG reduction levels) do not necessarily induce herd size reductions under the
advanced indicator, which lowers the impact of milk yield level variations on the MAC.
Independent of the GHG indicator chosen, variations of the stable age (building
year of the stables) as well as the time horizon show nearly the same eect: they slightly
reduce the MAC by increasing exibility with respect to the distribution of emissions
over time, which also allows for a better timing of abatement-related investments.
The wage rate, determining the opportunity costs for on-farm work, has a signicant but moderately negative impact on the MAC as higher returns to o-farm work
reduce the prot forgone from shrinking the farm operation.
Similarly, a lower labour productivity which increases the working hours spend per
cow and year (WHperCow) reduces the MAC for both indicators (the eect is slightly
smaller under the advanced indicator). Reducing GHGs with a low labour productivity releases more labour for o-farm work, which dampens income losses compared
with more labour ecient farms.
Higher concentrate prices let MAC decline somewhat under both indicators, which
is simply the opposite eect of an increased output price. The lower impact compared
with the milk price reects the cost share of concentrates.
5. Discussion of Results
Figure 4 plots the tted values for the MAC dierence in prots in relation to dierence in maximal emissions of the meta-models against those simulated with the
underlying complex bio-economic model (DAIRYDYN). The reader should rst note
the thick cluster of zero MAC observations along the horizontal axis, which are
rooted in basis changes in the MIP based simulation model DAIRYDYN. These
observations, hardly meaningful for economic analyses, are smoothed out by the
meta-models.
These zero observations contribute to the fact that we estimated a relatively low
adjusted R of about 0.46 and 0.47 for the meta-models of the simple and the
advanced indicators, respectively. Generally, it is hard to achieve a good t over the
non-smooth simulation behaviour of a LP or MIP with a regression model. The
reader should also be aware that, generally, the t of regression models for dierences
in dependent variables (here MAC) tend to be lower compared to level estimations
(here AC). Against this background, the t of the estimated meta-models can be valued as acceptable as it only consists of highly signicant regressors with a signicance
level of at least 5%.
As with any type of statistical model, results are conditioned on the input data and
hence only valid for the sampling background. Whereas the direction of impacts
2014 The Agricultural Economics Society
50
50
100
150
200
250
150
50
0
50 100
200
300
594
50
50
100
150
200
250
Figure 4. Scatter plot of tted values from meta-model against values simulated with
bio-economic model ( ton1 CO2-equ.)
found for the factors can be motivated from economic theory and thus generalised,
variable selection and related parameter estimates might depend on the sampling
design and are therefore probably only representative for dairy farm conditions in
North Rhine-Westphalia, Germany, as reected in the parameterisation and structure
of DAIRYDYN.
As simulation results were generated from specic assumptions about the length of
the simulation horizon and the currently observed production techniques implemented in the model structure of DAIRYDYN, an interesting question for further
research is whether and how to reect future technical progress in farming. While it is
obvious that innovations and their adaptation can change MAC (see, for example,
Amir et al., 2008), it is not clear how to account systematically for that eect in an
economic simulation with a longer time horizon.
Similarly, we should emphasise that single factor eects visualised in Figure 3 are
derived by xing all other factors to their median. Clearly, the single factor MAC
curves may change both with regard to their mean and slope when xing the other
factors to dierent levels. This may reorder the ranking of the factors with regard to
their impacts on MAC. The selected curves are thus only illustrative of the statistical
dependencies expressed by, and the simulations possible with the meta-models, while
underlining the usefulness of the meta-models for systematic analyses.
6. Summary and Conclusions
Our paper discusses the development of a statistical meta-model from simulations
with the highly detailed bio-economic single farm optimisation model DAIRYDYN,
in order to systematically analyse key factors impacting marginal abatement costs for
GHGs on dairy farms and to quantify their eects, in comparison between dierent
GHG indicators. DAIRYDYN was parameterised and simulations conducted to yield
a set of results covering the relevant range of core attributes representative of the
dairy farm population in North Rhine-Westphalia, Germany. In order to keep the
necessary simulations at a manageable size, we apply a design of experiments
approach, specically, a non-orthogonal Latin-Hypercube sampling approach which
accounts for correlations between factors. We observed farm exits for a non-negligible
share of simulations, which yield zero MAC for further emission enforcements.
2014 The Agricultural Economics Society
595
596
597
DeCara, S., Houze, M. and Jayet, P. A. Methane and nitrous oxide emissions from agriculture
in the EU: A spatial assessment of sources and abatement costs, Environmental and Resource
Economics, Vol. 32, (2005) pp. 551583.
Durandeau, S., Gabrielle, B., Godard, C., Jayet, P. and LeBas, C. Coupling biophysical and
micro-economic models to assess the eect of mitigation measures on greenhouse gas emission from agriculture, Climatic Change, Vol. 98, (2010) pp. 5173.
FDZ (Research Data Centre of the Statistical Departments of the Federation and the Federal
States, Germany) Census of Agriculture 2010, own calculations (Forschungsdatenzentrum
der Statistischen Landesamter, Kiel, Germany, 2013). Available at: http://www.forschungs
datenzentrum.de (last accessed 25 January 2013).
Fischer, C. and Morgenstern, D. Carbon abatement costs, why the wide range of estimates?
Discussion Paper 0342 (Resources for the Future, Washington, DC, 2005).
Flachowsky, G. and Brade, W. Potenziale zur Reduzierung der Methan-Emissionen bei
Wiederkauern, Z
uchtungskunde, Vol. 79, (2007) pp. 417465.
Giovanopoulou, E., Nastis, S. A. and Papanagiotou, E. Modeling farmer participation in agrienvironmental nitrate pollution reduction schemes, Ecological Economics, Vol. 70, (2011)
pp. 2,1752,180.
Giunta, A. A., Wojtkiewicz, S. F. and Eldred, M. S. Overview of modern design of experiments
methods for computational simulations, in 41st AIAA Aerospace Sciences Meeting and
Exhibit, AIAA 2003-0649 (Reno, NV, 2003).
Golub, A., Hertel, T., Lee, H. L., Rose, S. and Sohngen, B. The opportunity cost of land use
and the global potential for greenhouse gas mitigation in agriculture and forestry, Resource
and Energy Economics, Vol. 31, (2009) pp. 299319.
Heckman, J. J. Sample selection bias as specication error, Econometrica, Vol. 47, (1979)
pp. 153161.
Iman, R. L. Latin Hypercube sampling, in E. L. Melnick and B. S. Everitt (eds.), Encyclopedia
of Quantitative Risk Analysis and Assessment (England: John Wiley and Sons Ltd, 2008). doi:
10.1002/9780470061596.risk0299
Iman, R. L. and Conover, W. J. Small sample sensitivity analysis techniques for computer
models, with an application to risk assessment (with discussion), Communication in Statistics,
Vol. 9, (1980) pp. 1,7491,842.
Iman, R.L. and Conover, W.J. A distribution-free approach to inducing rank correlation
among input variables, Communications in Statistics Simulation and Computation, Vol. 11,
(1982) pp. 311334.
Iman, R. L., Campbell, J. E. and Helton, J. C. An approach to sensitivity analysis of computer
models. I - Introduction, input, variable selection and preliminary variable assessment, Journal of Quality Technology, Vol. 13, (1981) pp. 174183.
IPCC IPCC Guidelines for National Greenhouse Gas Inventories Prepared by the National
Greenhouse Gas Inventories Programme, H. S. Eggleston, L. Buendia, K. Miwa, T. Ngara,
and K. Tanabe (eds.), (Japan: IGES, 2006, pp. 10.111.54).
IT.NRW (Information und Technik Nordrhein-Westfalen) Statistische Berichte, Viehhaltung
und Viehbestande in Nordrhein-Westfalen am 1.M
arz 2010 nach Bestandsgr
oenklassen
(D
usseldorf, 2012). Available at: http://www.it.nrw.de (last assessed 13 February 2013).
Karplus, W. J. The spectrum of mathematical models, Perspective in Computing, Vol. 3, (1983)
pp. 413.
Kennedy, P. A Guide to Econometrics (Malden, MA: Blackwell Publishing, 2008).
Kleijnen, J. P. C. Verication and validation of simulation models, European Journal of Operational Research, Vol. 82, (1995) pp. 145162.
Kleijnen, J. P. C. Statistical validation of simulation, including case studies, in C. van
Dijkum, D. de Tombe and E. van Kuijk (eds.), Validation of simulation models, SISWO.
(Amsterdam: Netherlands Universities Institute for Coordination of Research in Social
Sciences, 1999).
598
599
Perez, I. Greenhouse gases: Inventories, abatement costs and markets for emission permits in
European agriculture, a modelling approach. (Germany, Frankfurt am Main: European
University Studies Vo.3184, Peter Lang Verlag GmbH, 2006).
Ramilan, T., Scrimgeour, F. G., Levy, G., Maarsh, D. and Romera, A. J. Simulation of
alternative dairy farm pollution abatement policies, Journal of Environmental Modelling and
Software, Vol. 26, (2011) pp. 27.
S
arndal, C. E., Swensson, B. and Wretman, J. Model Assisted Survey Sampling (New York:
Springer-Verlag, 1992).
Scheele, M., Isermeyer, F. and Schmitt, G. Umweltpolitische Strategien zur L
osung der Stickstoproblematik in der Landwirtschaft, Agrarwirtschaft, Vol. 42, (1993) pp. 294313.
Schneider, U. A. and McCarl, B. A. Appraising agricultural greenhouse gas mitigation
potentials: Eects of alternative assumptions, Agricultural Economics, Vol. 35, (2006)
pp. 277287.
Smith, P., Martino, D., Cai, Z., Gwary, D., Janzen, H., Kumar, P., McCarl, B., Ogle, S.,
OMara, F., Rice, C., Scholes, B., Sirotenko, O., Howden, M., McAllister, T., Pan, G.,
Romanenkov, V., Schneider, U., Towprayoon, S., Wattenbach, M. and Smith, J. Greenhouse gas mitigation in agriculture, Philosophical Transactions of the Royal Society Biological Sciences, Vol. 363, (2008) pp. 789813.
Stoker, T. M. Empirical approaches to the problem of aggregation over individuals, Journal of
Economic Literature, Vol. 16, (1993) pp. 1,8271,874.
Tobin, J. Estimation of relationships for limited dependent variables, Econometrica, Vol. 26,
(1958) pp. 2436.
Toomet, O. and Henningsen, A. Sample selection models in R: Package sample Selection,
Journal of Statistical Software, Vol. 27, (2008) pp. 123.
Vermont, B. and DeCara, S. How costly is mitigation of non-CO2 greenhouse gas emissions
from agriculture? A meta-analysis, Ecological Economics, Vol. 69, (2010) pp. 13731386.
Weiske, A. and Michel, J. Greenhouse gas emissions and mitigation costs of selected mitigation
measures in agricultural production, Final Version 08.01.2007. Specic targeted research
project no. SSPE-CT-2004-503604 (MEACAP WP3 D15a, 2007).
Wyss, G. D. and Jorgensen, K. H. A Userss Guide to LHS: Sandias Latin Hypercube Sampling
Software. SAND98-0210 (Albuquerque: NM Sandia National Laboratories, 1998).