Analysis of farm performance in Europe under different
climatic and management conditions to improve
understanding of adaptive capacity
Pytrik Reidsma & Frank Ewert & Alfons Oude Lansink
Received: 7 June 2005 /Accepted: 10 January 2007 / Published online: 7 March 2007
# Springer Science + Business Media B.V. 2007
Abstract The aim of this paper is to improve understanding of the adaptive capacity of
European agriculture to climate change. Extensive data on farm characteristics of individual
farms from the Farm Accountancy Data Network (FADN) have been combined with climatic
andsocio-economicdatatoanalyzetheinfluenceofclimateand managementon cropyieldsand
income and to identify factors that determine adaptive capacity. A multilevel analysis was
of climate conditions on farm yields and income. Next to climate, input intensity, economic size
yields and income. Generally, crop yields and income are increasing with farm size and farm
intensity. However, effects differed among crops and high crop yields were not always related to
high incomes, suggesting that impacts of climate and management differ by impact variable. As
farm characteristics influence climate impacts on crop yields and income, they are good
indicators of adaptive capacity at farm level and should be considered in impact assessment
models. Different farm types with different management strategies will adapt differently.
Climate change is expected to affect agriculture very differently in different parts of the
world (Parry et al. 2004). Many studies have analyzed the influence of climate and climate
Climatic Change (2007) 84:403–422
P. Reidsma (*):F. Ewert
Department of Plant Sciences, Group Plant Production Systems, Wageningen University,
P.O. Box 430, 6700 AK Wageningen, The Netherlands
Netherlands Environmental Assessment Agency (RIVM/MNP),
P.O. Box 1, 3720 BA Bilthoven, The Netherlands
A. Oude Lansink
Department of Social Sciences, Business Economics, Wageningen University,
P.O. Box 8130, 6700 EW Wageningen, The Netherlands
change on agriculture, and the problem of agricultural vulnerability is increasingly
recognized (e.g. Mendelsohn et al. 1994; Antle et al. 2004; Parry et al. 2004). The extent
to which systems are vulnerable depends on the actual exposure to climate change, their
sensitivity and their adaptive capacity (IPCC 2001). Exposure and sensitivity determine the
potential impacts, which include all impacts that occur given the projected climate change
without considering adaptation. The actual impact is the impact that remains after allowing
for adaptation. The adaptive capacity refers to the ability to cope with climate change
including climate variability and extremes in order to (a) moderate potential damages,
(b) take advantage of emerging opportunities, and/or (c) cope with its consequences. Most
quantitative studies that address the vulnerability of agricultural systems have focussed on
exposure and sensitivity, while adaptive capacity is often highly simplified. Realistic
adaptation processes are not well understood and therefore hard to quantify (Smit et al.
The impact of climate change on society is frequently determined by assessing impacts
on ecosystem services (Metzger 2005; Reid et al. 2005). Because ecosystem services form a
direct link between ecosystems and society, the concept is especially useful for illustrating
the need to employ mitigation or adaptation measures to prevent or alleviate impacts
(Metzger 2005). The main ecosystem services provided by the agricultural sector are food
production, farmers’ income and environmental sustainability. Impacts of climate change
on food production are generally assessed with crop models (Gitay et al. 2001). Studies
have been performed on different levels of organization: crops (Tubiello and Ewert 2002),
cropping systems (e.g. Tubiello et al. 2000), regional (Iglesias et al. 2000; Saarikko 2000;
Trnka et al. 2004), continental (Harrison et al. 1995; Downing et al. 2000; Reilly 2002) and
global (IMAGE Team 2001; Parry et al. 2004).
In crop modelling studies, farmers’ responses to climate change are purely hypothetical
and either no adaptation or optimal adaptation is assumed. Easterling et al. (2003) made a
first attempt to model agronomic adaptation more realistically proposing a logistic growth
function to describe the adaptation process over time. How agricultural adaptive capacity
varies spatially has not been assessed to date, however. Mendelsohn and Dinar (1999)
suggest that climatic conditions have relatively smaller impact on farmers’ income (net
income/farm value) than on crop yields as simulated by crop models. Their cross-sectional
analysis implicitly includes adaptive capacity. Adaptation strategies adopted could be
agronomic strategies to increase crop yields as well as economic strategies such as changes
in crops and inputs. Agro-economic models (Kaiser et al. 1993; Antle et al. 2004) can
assess optimal economic adaptation strategies, but do not consider the capacity to adapt
these. In addition, biophysical relationships are often underrepresented.
In Europe, concerns in agriculture are mainly related to farmer livelihood and the land
available for farming (Schröter et al. 2005) and less to food production. A European
vulnerability assessment showed that farmer livelihood is especially vulnerable in the
Mediterranean region (Metzger et al. 2006). This projection was based on calculations
suggesting that intensification of production will reduce the need for agricultural land in
less favoured areas (Ewert et al. 2005; Rounsevell et al. 2005). Although the impact of
climate change in Europe was projected to be small on average, regions with less
favourable climatic conditions and hence lower crop yields would have difficulties to
sustain farmer livelihood. Projected impacts on European agricultural land use were less
severe when the global food market and regional land supply curves were included in the
modelling framework (van Meijl et al. 2006). Assumptions related to different drivers have
a large influence on climate change impact projections. Farm-level responses are usually
404 Climatic Change (2007) 84:403–422
not considered and spatial variability in farm performance and adaptive capacity is not well
In this paper we analyzed the impact of farm characteristics and climatic and socio-
economic conditions on crop yields and farmers’ income across the EU15. The influence of
climate is assessed using a Ricardian approach, similar to that employed by Mendelsohn et al.
(1994). By including farm-level information (e.g. farm size, intensity) and socio-economic
conditions in the analysis, we captured factors that influence farm-level adaptive capacity.
We investigated both crop yields and income variables and the relationships between these
to understand farm performance and adaptation.
Emphasis is on spatial variability in farm performance considering data from three
different years (1990, 1995 and 2000). Since data were available at different scales a
multilevel statistical approach was used. Results of this study can improve the modelling of
agricultural adaptation to climate change.
2.1 Conceptual basis for analyzing farm performance and adaptive capacity
Changes in climatic conditions will affect crop growth and yield at the field level through
biophysical relationships and these impacts are commonly assessed with crop models. The
dynamic nature of climate effects is well understood for potential, water and nitrogen
limited growth and yield (e.g. van Ittersum et al. 2003). Actual yields, however, are also
affected by other factors such as pests and diseases not considered in crop models and farm
management will largely influence the obtained actual yield. Therefore, climate change
impacts on crop yields also depend on factors determining farm performance. Potential
impacts can be assessed with crop models, but for projections of actual impacts the adaptive
capacity of farmers should be taken into account.
We found it important to distinguish between two groups of factors related to (1) farm
characteristics and (2) regional conditions such as biophysical, socio-economic and policy
factors (Fig. 1). Both factor groups represent different levels of organization (farm and
region). We account for possible interactions between farm characteristics and regional
conditions on farm performance through a multilevel analysis (see Section 2.3). Farm
characteristics may also change as a result of regional impacts on farm performance, which,
however, is not further addressed in this paper. As different crops respond differently to
Regional conditions Regional conditions
(climate, soil, …) (climate, soil, …)
(welfare, technology, prices,…) (welfare, technology, prices,…)
(subsidies, regulations,…)(subsidies, regulations,…)
Farm(er) characteristics Farm(er) characteristics
(ecosystem services) (ecosystem services)
Crop yields Crop yields
Farmers’ incomeFarmers’ income
Fig. 1 The investigated relation-
ships (represented by the block
arrows). Potential impacts of cli-
mate conditions are influenced by
other regional conditions and
farm characteristics, which deter-
mine adaptive capacity
Climatic Change (2007) 84:403–422405
climatic conditions, yields of five important crops (wheat, grain maize, barley, potato and
sugar beet), were analyzed.
Farm management decisions have to be economically viable in order to ensure the farm’s
sustainability. We considered the economic performance of farms by including farmers’
income in the analysis and explicitly studied relationships between income and crop yields.
Farmers’ income is represented by farm net value added per hectare (fnv/ha) and farm net
value added/annual work unit (fnv/awu). Fnv/ha measures economic performance per unit
of land and a relationship to crop yield can be expected. Fnv/awu is a measure that enables
comparison of farmers’ income directly to GDP per capita and can therefore relate farm
performance to general socio-economic performance. By directly measuring revenues, we
account for the direct impacts of climate on yields of different crops as well as the indirect
substitution of different inputs, introduction of different activities, and other potential
adaptations to different climates (Mendelsohn et al. 1994).
Farm characteristics that explain farm performance are related to determinants of
adaptive capacity: awareness, technological ability and financial ability (Schröter et al.
2003; Metzger et al. 2006). Adaptive capacity is difficult to quantify explicitly from
observations on farm performance however. Information about potential impacts, i.e.
impact without adaptation, is not available as observed farm performance implicitly
includes adaptation to present climatic and other conditions. We assume that adaptation is
related to farm performance and farms that perform well are also well adapted.
2.2 Data sources and data processing
The Farm Accountancy Data Network (source: FADN-CCE-DG Agri and LEI) provides
extensive data on farm characteristics of individual farms throughout the EU151. Data have
been collected annually since 1989. They have been used as an instrument to evaluate the
income of agricultural holdings and the impacts of the Common Agricultural Policy.
Information about the exact geographic location of the sample farms is not available for
privacy reasons; only the region in which farms are located is known. In total, 100 HARM
regions2are distinguished (see Fig. 3) with 51,843 sample farms.
FADN considers the following land-using production types: specialist field crops,
specialist permanent crops, specialist grazing livestock, mixed cropping and mixed crops/
livestock. At approximately 40% of all farms, i.e. 20,936 farms, crop production is the main
activity, i.e. when more than 66% of the total standard gross margin3(economic size) was
obtained from the sale of field crop products and/or when the arable area was more than
66% of the total utilized agricultural area. Only these farms were included in the analysis of
effects on farmers’ income.
For each farm, data were available on outputs representing farm performance: crop
yields and farm net valued added. Crop yields of five important crops (wheat, grain maize,
barley, potato and sugar beet) were calculated by dividing production (in tons fresh matter)
by crop area (in ha). Farm characteristics considered to explain farm performance represent
different determinants of adaptive capacity: awareness, technological ability and financial
1The EU15 comprises the 15 member countries of the European Union before the extension in 2004.
2HARM is the abbreviation for the harmonized division created by the Dutch Agricultural Economics
Research Institute (LEI). It gives the opportunity to compare the different regional divisions of the EU15
used by Eurostat (NUTS2) and FADN.
3The standard Gross Margin (SGM) of a crop or livestock item is defined as the value of output from one
hectare or from one animal less the cost of variable inputs required to produce that output.
406 Climatic Change (2007) 84:403–422
ability (Schröter et al. 2003; Metzger et al. 2006). Awareness is reflected in the land use
(arable land, permanent cropping land, grassland, area of each crop grown). Arable farmers
have more skills in crop production than livestock farmers and therefore obtain higher
yields and probably less yield variability. A farmer growing a specific crop in a large area is
expected to put more effort in obtaining a high crop yield. Technological ability is
represented by the input intensity (irrigated area, input costs of fertilizer and crop protection
products, whether the farm is conventional or organic). It is expected that farms with a high
input intensity aim for a high output intensity. Financial ability is reflected by the economic
size and/or the size of the farm in hectares. A larger farm is a priori expected to have more
capital available for investments in new technologies. Altitude class and location in a less-
favoured area (LFA) were used as proxies for the biophysical characteristics of the land.
More variables were available, but variables needed to be selected to reduce multi-
collinearity (see Sections 2.3.2 and 3.2). Data from three years (1990, 1995 and 2000) were
considered but results presented refer mainly to the year 2000 as little or no differences
were found among years.
Climatic effects were analyzed using data from the ATEAM project4based on New et al.
(2002). Averages from the 30-year period 1971–2000 are assumed to be representative for
the climatic conditions that influence spatial variability in farm performance.5Mean
temperature and precipitation of all months were obtained with a resolution of 100×100. As
monthly climate variables are often correlated, average variables were created to not
confound the results. Monthly mean temperatures of the first six months (January–June)
have been averaged, resulting in the mean monthly temperature of the first half of the year.
Also precipitation data was averaged to obtain the mean monthly precipitation for the first
6 months of the year that can be considered as the main growing period for Europe. All
climatic data were averaged to HARM regions.
Data on regional socio-economic variables, such as GDP per capita and population
density were obtained from Eurostat (2004). Population density can serve as a proxy for the
pressure on the land. When land becomes scarce, rental rates increase, which is assumed to
increase production intensity (Van Meijl et al. 2006). Data were available at NUTS26level
and transformed to HARM regions.
A macro-scale adaptive capacity index has been developed at NUTS2 regional level for
the EU15 (Schröter et al. 2003; Metzger et al. 2006). This adaptive capacity index serves as
a proxy for the socio-economic conditions that influence farmers’ decisions; it sets the
regional context in which individuals adapt. The index is based on twelve indicators, which
are aggregated by application of fuzzy set theory. The indicators comprise: female activity
rate & income inequality (equality), literacy rate & enrolment ratio (knowledge), R&D
expenditure & number of patents (technology), number of telephone lines & number of
doctors (infrastructure), GDP per capita & age dependency ratio (flexibility), world trade
share & budget surplus (economic power). In Table 1 a description is given of all variables
used in the analysis.
6Nomenclature des Units Territoriales Statistiques 2: regions or provinces within a country as distinguished
5Spatial variability in crop yields and income is mainly determined by long-term climate variability.
Temporally, variability in crop yields and income is relatively smaller than climate variability (results not
shown). Using yearly climate data disturbs the impact of long-term spatial variability in climatic conditions.
4ATEAM (Advanced Terrestrial Ecosystem Analysis and Modelling), http://www.pik-potsdam.de/ateam/
Climatic Change (2007) 84:403–422407
2.3 Statistical analysis
2.3.1 Multilevel modelling
The effect of climate and management on farm performance is analyzed by fitting a
multilevel (or generalized linear mixed model; GLMM) model to the data. A multilevel
model expands the general linear model (GLM) so that the data are permitted to exhibit
correlated and non-constant variability (e.g. Snijders and Bosker 1999; McCulloch and
Searle 2001). Multilevel modelling originates from the social sciences and has more
recently also been applied to geographic studies (e.g. Polsky and Easterling 2001; Pan et al.
2004). A multilevel model can handle complex situations in which experimental units are
Table 1 Data description and sources
Actual crop yield (tons/ha)
Farm net value addedd/annual work units (€)
Farm net value added/hectare (€)
Irrigated percentage of utilized agricultural area (%)
Costs of fertilizers and soil improvers per hectare (€)
Costs of crop protection products per hectare (€)
1=conventional, 2=organic, 3=converting/partially
Utilized agricultural area (ha)
Annual work units (AWUf)
Permanent cropping area/utilized agricultural area (–)
Grass/uaa* Grassland area/utilized agricultural area (–)
Crop_pr* Crop area/total arable area (–)
Alt* Altitude: 1=<300 m, 2=300–600 m, 3=>600 m
Lfa*1=not in lfag, 2=in lfa not mountain, 3=in lfa mountain 1
Tmean*Mean monthly temperature (°C) of first half year
Pmean*Mean monthly precipitation (mm) of first half year
Ac* Macro-scale adaptive capacity index (–)
Gdp/cap Gross domestic product per capita (€)
Pop_dens Population density (people per km2)
*Independent variables included in multilevel models
a1: FADN, 2: ATEAM, 3: Eurostat (1=farm level; 2,3=HARM level).
bStatistics based on 2000 data, for cropping systems only.
cDiffers per crop considered.
dCorresponds to the payment for fixed factors of production (land, labour and capital), whether they are
external or family factors. As a result, holdings can be compared irrespective of the family/non-family nature
of the factors of production employed. Fnv=total output−total intermediate consumption+balance current
subsidies and taxes−depreciation.
eThe economic size is determined on the basis of the overall standard gross margin of the holding. It is given
in European Size Units (ESU); one ESU corresponds to a standard gross margin of €1,200.
fOne Annual Work Unit (AWU) is equivalent to one person working full-time on the holding.
gLfa = Less-favoured area.
408 Climatic Change (2007) 84:403–422
nested in a hierarchy. In a multilevel model, responses from a subject are thought to be the
sum of the so-called fixed and random effects. If a variable, such as fertilizer use, affects
wheat yield, it is fixed. Random effects contribute only to the covariance of the data.
Intercepts and slopes of variables may vary per region and this covariance is modelled
using random effects. Hence, multi-level modelling accounts for regional differences when
analyzing within region effects of farm characteristics on yields and income. In Fig. 2 this is
Fitting a multilevel model to the data comprises a few steps. Firstly, the model is
formulated with fixed effects only as in a GLM, to compare against models including
different forms of HARM-level variation.
In Eq. 1, yijis the dependent variable, b0jis the intercept estimate, bqjis the coefficient
estimate of the variable xqj, i indexes the farm, j indexes the HARM region and the residual
rij∼N(0, σ2). In this model, b0jand bqjare the same for all HARM regions. The model gives
similar results as a GLM. The goodness of fit is measured in different ways though. A
multilevel model is based on (restricted) maximum likelihood methods, versus the
minimization of squared error in GLM. The preferred GLM is the model with the highest
R2, while the preferred multilevel model is selected using likelihood ratio tests. The preferred
multilevel model is the model with the lowest information criteria, such as −2 log likelihood
(deviance) or Aikaike’s Information Criterion (AIC). A single deviance or AIC has no useful
interpretation, it is only the difference between the values of different models that matters.
In a second model, the proposition that the average of the dependent variable varies
between regions is being tested by including a random intercept. This model combines
Eqs. 1 and 2.
b0j¼ b0þ mj
where μjis the regional level residual from the average intercept estimate. To test whether
the overall model fit is improved, two models can be compared by subtracting the
deviances. This is the χ2, and the associated d.f. is the difference in the number of
Fertilizer useFertilizer useFertilizer use Fertilizer use Fertilizer useFertilizer use
Fig. 2 Graphicalexampleofamultilevelmodelwitha random intercept b0jand b random intercept b0jand slopes bqj.
Each solid line represents the effect of fertilizer use on wheat yield in a specific region j, whilst the dotted line
represents the mean (fixed) relationship across all regions (bq0). In a simple regression model, the mean relationship is
a line through all the data points, while in a multilevel model it’s the average of the relationships per region. See
Section 2.3.1 for further explanation
Climatic Change (2007) 84:403–422409
parameters. A random intercept model allows for a better representation of the influence of
farm-level variables on the dependent variables, as regional differences are being captured
in the random intercept. Since the focus is on the explanation of variables within regions,
regional differences in climatic or socio-economic conditions which are not captured by the
selected variables, do not confound the results. The influence of variables can also differ
between regions. We therefore tested the random coefficients model, in which also the
slopes vary between regions. This model combines Eqs. 1, 2 and 3.
bqj¼ bq0þ uqj
where uqjis the regional level residual from the average coefficient estimate. All statistical
analyses were performed with the data of the years 1990, 1995 and 2000 separately. Since
results were consistent across years only results from 2000 are presented (see Section 3).
2.3.2 Selection of variables
Crop yields (wheat, grain maize, barley, potato and sugar beet) and income variables (farm
net value added/annual work unit, farm net value added/ha) were the dependent variables in
different models. These and the independent variables are presented in Table 1. For the
climate variables, linear and quadratic terms were included to capture their potential
nonlinear effects on crop yields and income variables. For crop yield models all sample
farms in the database were analyzed, for income models only farms where crop production
was dominating were considered (see Section 2.2).
The two-way relationship between the dependent variables and fertilizer and crop
protection use violates a basic assumption of independence and therefore can lead to
endogeneity. Farmers’ decisions about the rate of fertilizer and crop protection applications
depend on its marginal effects on the net value added, which is determined by the marginal
effect on crop yields, the prices of crops, and the prices of fertilizers and crop protection
products. Non-linearity of the relationship between these input costs and dependent
variables has been tested by curve estimation in SPSS 11. To test for the impact of
erroneously treating endogenous variables as exogenous, we used instrumental variables
(IV) to estimate the effect of fert/ha and prot/ha on the dependent variables. Using
instrumental variables allows for removing the error terms in fert/ha and prot/ha that
confound with the errors in the equations of crop yields and farm income. All variables in
the database that could possibly influence application of fert/ha and prot/ha were included
as instrumental variables in the IV regression (e.g. land improvement costs, costs on
machinery and equipment, percentages of various crops, annual working units7). The IV
regression was performed with a multilevel model. Endogeneity of fert/ha and prot/ha was
tested by the Hausman test (Hausman 1978). The test statistic is
whereeb is the parameter vector resulting from the model based on IV estimates for the
values.eV andbV are the variance-covariance matrices ofeb andbb, respectively. This test has
M ¼ eb ?bb
possible endogenous variables andbb is the parameter vector of the model with the observed
a χ2distribution with N degrees of freedom (N is the number of parameters). The null
7A full list of variables used in the instrumental variables regression can be obtained from the corresponding
410 Climatic Change (2007) 84:403–422
hypothesis is that the two estimators do not differ. If the null hypothesis is rejected,
exogeneity of the variables under investigation is rejected. The Hausman test can result in
negative test values. One way to deal with this is to apply the test on the parameters tested
for endogeneity only (Ooms and Peerlings 2005).
Climate, socio-economic and management variables all have, to some extent, a north–south
gradient in the European Union. A high multicollinearity causes coefficient estimates to be
unreliable and confounding in interpreting the model results. An advantage of a full multilevel
the influence of management variables is analyzed per region (as random effects account for
regional differences), a possible correlation of input use (at individual farm level) with climatic
variables (at regional level) won’t influence the results.
The linear mixed model procedure in SPSS 11 does not include collinearity diagnostics.
We therefore applied a linear regression model to the data to examine these. We based the
selection of variables on the partial correlation matrix and on the linear regression model
with wheat yield as dependent variable. Firstly insignificant variables were removed;
secondly variables with a variance inflation factor (VIF) of 10 or higher were removed from
the analysis (Allison 1999). The process of excluding variables was continued until all
condition indices (CI) were below 30 and all variables contributed to the output. CI greater
than 30 indicate that multicollinearity is a serious concern; multicollinearity is not present
when all condition indices equal one.
3.1 Spatial variability in yield and income variables
InFig. 3 the spatial variability of wheat yield, maize yield, farm net value added/annual work
unit (fnv/awu) and farm net value added/hectare (fnv/ha) between and within HARM regions
in 2000 is presented. The coefficient of variation (CV) gives an indication of the spatial
variability within a region due to management and/or biophysical factors. Spatial distributions
of yields were different for wheat and maize. Wheat yields were generally highest in
northwest Europe, while the highest maize yields were obtained in Spain and Greece. Spatial
variability within regions was generally higher in regions with lower yields. The variability
among regions of fnv/awu was similar to that of wheat yields, but different to the spatial
variability of fnv/ha which was especially high for some Mediterranean regions.
3.2 Selection of variables affecting crop yield and income
The instrumental variables regression model could account for 81.2% of the variation in
fert/ha and 83.1% of prot/ha. Results of the Hausman test indicated that fertilizer use and
crop protection use were exogenous to crop yields (p>0.05), but endogenous to fnv/ha and
fnv/awu (p<0.001). Hence the observed values were used in the crop yield models, while
the estimates based on the IV model were used in the income models.
In a partial correlation matrix (Table 2) we identified variables that were correlated, and
variables that were correlated to the dependent variables in which we were interested. The
correlation between crop protection use (prot/ha) and wheat yield for example was
significantly positive with an r2=0.467, suggesting that prot/ha may be a good predictor of
wheat yield and should be included in the multilevel model.
Climatic Change (2007) 84:403–422411
For each model it was tested whether including quadratic terms improved model
performance. Models that include mean temperature (tmean), as well as the macro-scale
adaptive capacity (ac) showed Variance Inflation Factors of nearly 2 and Condition Indices
higher than 30, which indicates that coefficient estimates were not reliable. For each model
Fig. 3 Spatial variability of crop yields (tons/ha) and income variables (€) in 2000 between and within
HARM regions for a average wheat yield, b CVof wheat yield, c average maize yield, d CVof maize yield,
e average of farm net value added/annual work unit (fnv/awu), f CVof fnv/awu, g average of farm net value
added/hectare (fnv/awu) and h CV of fnv/ha. Only values for regions where more than 15 farms grow the
crop considered are presented
412Climatic Change (2007) 84:403–422
Table 2 Partial correlation matrix of selected variables in 2000 for farms with crop production as the main farming activity
0.209 −0.003 −0.028
perm/uaa −0.183 −0.064 −0.142 −0.110
0.031 −0.131 −0.137
0.018 −0.027 −0.105 −0.117
−0.106 −0.125 −0.117
0.106 −0.081 −0.070
−0.074 −0.040 −0.094
0.134 −0.041 −0.038
−0.043 −0.038 −0.005 −0.084 −0.013
−0.196 −0.076 −0.012
0.040 −0.133 −0.080
0.231 −0.145 −0.144 −0.160 −0.079
−0.122 −0.013 −0.260
0.018 −0.008 −0.041
0.082 −0.129 −0.103 −0.062 −0.104
0.105 −0.150 −0.422
−0.112 −0.021 −0.183
0.060 −0.687 0.120
0.099 −0.097 −0.334
0.009 −0.081 −0.621 0.203
0.026 −0.010 −0.025 0.123
Pearson’s correlation coefficients (r2) in bold are significant. Names of crops refer to actual yields. Other variables are described in Section 2.2 and Table 1.
Climatic Change (2007) 84:403–422413
either climate variables or the ac have been included. Gdp/cap was highly correlated with
ac and was excluded from further analysis. Both variables can represent the socio-economic
conditions influencing farmers’ decision making; however, ac is more comprehensive and a
better indicator of the regional context in which individuals adapt. Although population
density (pop_dens) had a significant positive effect on wheat and maize yields and fnv/awu,
its effect was not significant in multilevel models and was excluded from further analysis.
On the individual farm level, the size of the farm in hectares (uaa) and labour units
(labour) were highly correlated with the economic size of the farm (ec_size). Only ec_size
was included in the multilevel models. As the share of arable land (ar/uaa), permanent
cropping land (perm/uaa) and grassland (grass/uaa) in total uaa almost add up to one, they
can not all be included in the model. Consequently, ar/uaa is excluded from the model.
Thus, a negative effect of the other land use types implies a positive effect of ar/uaa.
3.3 The influence of climate and management on crop yields
The multilevel model with wheat yield as dependent variable clearly improved when
random intercepts and slopes were introduced. The deviance decreased from 61,744 for a
model with fixed effects only, to 57,104 (p<0.001) when a random intercept was included,
to 55,735 (p<0.001) when random slopes were included. The covariance parameters of the
random effects were significant for all variables, indicating significance of between-region
variation. Thus, for estimating parameters of fixed effects it is better to use the model with
random intercept and slopes; this also holds for all other crop yield models.
Table 3 presents the fixed effects of multilevel models with random intercept and slopes.
The coefficient estimates refer to models with climate variables included. However, since
we were also interested in the effects of ac, coefficient estimates for ac (i.e. without climate
variables) are shown.
Wheat yield was significantly related to all variables included in the model, except for
irrigated percentage (irr_perc). The parameter estimates of the linear and quadratic terms of
mean temperature (tmean) and precipitation (pmean) suggests that relationships with wheat
yield were concave in these variables. Variables representing input intensity (fertilizer use,
fert/ha; crop protection use, prot/ha; conventional/organic farming, org) and financial
ability (economic size, ec_size) all influenced wheat yields significantly positive. The type
of land use also influenced wheat yield significantly: the percentage of wheat area
(crop_pr) had a positive effect and the percentage of permanent cropping area (perm/uaa)
and grassland area (grass/uaa) had a negative effect, indicating a positive effect for the
percentage of arable land (ar/uaa). The influence of irr_perc was not significant, which was
probably due to the fact that wheat is usually not irrigated. Effects of factors representing
growing conditions were highly significant. Farms on higher altitudes (alt) and farms in less
favoured areas (lfa) had, ceteris paribus, lower wheat yields compared to farms under more
favourable conditions. These results suggest that climatic conditions influence wheat yields,
but that farm characteristics can increase or diminish this influence.
Relationships for maize yields were less clear than for wheat. Effects of tmean were only
significant at p<0.10, while the effect of pmean was not significant. Variation in pmean
across Europe was relatively small and availability of water depends also on other factors
such as soil water holding capacity and depth and potential evapo-transpiration. In regions
with a low water availability irrigation is applied to maize.
Including quadratic terms of climate variables didn’t improve model performance
(in terms of AIC). For some farm characteristics such as irr_perc, fert/ha and perm/uaa
significant effects were evident. The maize growing area (crop_pr) was significant at p<0.10,
414 Climatic Change (2007) 84:403–422
Table 3 Fixed effects of multilevel models of 2000 with random intercept and slopes, with crop yields and income variables as dependent variables
***p<0.001; **p<0.01; *p<0.05;†p<0.10.
aSignificant (p<0.05) in models with fixed effects only. If the sign changes, this is indicated between brackets.
bSignificant (p<0.05) in models with fixed effects and random intercept. If the sign changes, this is indicated between brackets.
c‘Crop’ is the crop concerned in the column.
dThe models presented exclude the adaptive capacity index ac. When the ac is included instead of the climate variables, these are the coefficient estimates and significance
levels. Coefficient estimates of other variables are similar.
Climatic Change (2007) 84:403–422 415
but highly significant in models with fixed effects only, suggesting that maize yields were,
ceteris paribus, higher in regions were more maize was grown. Effects on yield were also
observed for ec_size but were only significantly positive in a model without random slopes.
This means that within regions, farms with a large economic size generally obtain higher
maize yields. In models with random slopes other variables can account for this however. The
negative effect in the fixed effects model suggests higher yields in regions with mainly
smaller farms. The correlation between prot/ha and maize yield (Table 2) was not confirmed
in the multilevel model. Maize yields were lower on organic farms (org), at higher altitudes
(alt) and in less favoured areas (lfa).
Results for barley were similar to the ones for wheat for most variables which was also true
for potato and sugar beet. Although these root crops are often irrigated, there was no
significant relationship between irr_perc and yield. This result is explained by the fact that in
regions with insufficient precipitation these crops are always irrigated, whereas in regions
with sufficient precipitation no irrigation takes place. Hence, variation among farms is
insufficient to identify a significant effect. Tmean had a non-linear influence on barley, potato
and sugar beet yields, whereas the influence of pmean was not significant. The effect of ac on
crop yield was positive for all crops, although not always significant in models with random
effects. This suggests some influence of the regional context for farm-level adaptation.
3.4 The influence of climate and management on income variables
3.4.1 Variability in farmers’ income
Multilevel models with farm net value added/annual work unit (fnv/awu) and farm net value
added/hectare (fnv/ha) as dependent variable, clearly improved with random intercept and
slopes. Applying a random coefficients model to the data can thus give better insight in the
effect of specific variables on farmers’ income. Fnv/awu was significantly positive related
to ec_size and ac and negative to fert/ha, perm/uaa and grass/uaa. The relation with tmean
was concave; there was no significant relation with pmean. For fnv/ha, effects of fert/ha and
prot/ha were significantly positive. Although not always significant, organic farming,
altitude and a less favoured area location generally had a positive effect on fnv/ha, whereas
they had a negative effect on fnv/awu.
The positive effect of variables representing input intensity on fnv/ha was not evident for
fnv/awu. On the other hand, variables that did not influence fnv/ha, like ec_size and ac, had
an effect on fnv/awu. Results show that intensification leads to higher fnv/ha, but also that
fnv/awu is, ceteris paribus, higher on larger farms and on farms with a lower intensity.
Enlargement thus seems to be a better adaptation strategy than intensification. However, it
is evident that farmers’ income is influenced by most farm characteristics considered.
Fnv/ha was not related to climate variables, whereas tmean had a non-linear concave effect
and pmean a negative effect on fnv/awu. This was surprising, as especially fnv/ha, which should
reflect the productivity of the land, was expected to be influenced by climatic variables. Ap-
parently, the relationship between crop productivity and farmers’ income is not straightforward,
as also evident from the change in signs in models without random effects and the (non-
significant) negative effect of ac on fnv/ha, which was positive for crop yields and fnv/awu.
3.4.2 Relationship between crop yields and farmers’ income
There was a highly significant relationship at the regional level between yields of most
crops and fnv/awu [wheat, r2=0.685; barley, r2=0.638; sugar beet, r2=0.407; potato, r2=
416Climatic Change (2007) 84:403–422
0.348; maize, r2=0.209 (only significant at the p<0.10 level)]. These correlations were also
significant at the farm level, but less pronounced (Table 2). Although a causal relation can
be assumed, this relation seems to be confounded by other factors. Income was highly
distorted by government support programs; the highest subsidies were received in the same
regions where the highest wheat yields were observed (e.g. northern France, England, East
Germany). Fnv represents the sum of revenues from outputs (O) − variable input costs (I) +
subsidies − taxes. The average O – I was negative in these regions, but due to subsidies the
average fnv became positive. Although average fnv/ha was still low, the large farm sizes
resulted in high fnv/awu.
Thus, fnv/ha was not related to crop yields and was especially high in many
Mediterranean countries with typically lower crop yields and smaller farms (note, however,
that Table 2 shows a small positive within region correlation between fnv/ha and yields of
some crops). This suggests that maximizing crop yields is not always an efficient economic
strategy. Clearly, differences in fnv/awu in Europe were mainly determined by farm size and
subsidies, while climatic conditions played a minor role.
3.5 Separating between climatic and management effects
Results from a multilevel analysis cannot directly differentiate between climate and
management effects. However, the influence of farm characteristics can be identified by
comparing the influence of tmean estimated by a multilevel model including climatic
conditions and farm characteristics with the influence estimated by a model only including
climatic conditions (Kaufmann and Snell 1997). An example is provided for wheat yield
(Fig. 4a). Omitted-variable bias in the model only including climatic variables causes
overestimation of the direct effect of tmean, as the effect of farm characteristics is forced
into the parameter estimates of the climatic variables. As a result, the reduction in yield
when climate conditions move away from the optimum are much more severe in the model
including only climate variables compared to the model with all variables included. This
suggests that current wheat management in relation to the variables included in the model
amplifies the effect of climatic conditions in less favourable areas. The exacerbated climate
effect in less favourable areas can be explained by (1) less- favourable socio-economic
68 10 1214
maize yield (tons/ha)
model with all variables
model including only climate variables
-4 -202468 10 12 14
wheat yield (tons/ha)
model with all variables
model including only climate variables
Fig. 4 The effect of tmean (°C) on a wheat yield (tons/ha) and b maize yield (tons/ha), based on the full
multilevel model, including climate variables and farm characteristics (thick line) and a model only including
climate variables. A value of zero represents no reduction in yield and is the physiological optimum for that
variable. The model only including climate variables indicates the total impact of tmean, while the full
multilevel model indicates the impact that can directly attributed to tmean. The difference between both lines
indicates the amplifying effect of farm characteristics on the impact of tmean on crop yields
Climatic Change (2007) 84:403–422417