Table of Contents
21st International System Dynamics Conference
New York, July 20-24, 2003
Policy and outcome contrasts in the evaluation of the
effects of structural change in Swiss mountain agriculture
using Linear Programming and System Dynamics
Birgit Kopainsky, Christian Flury, Peter Rieder
Agricultural Economics, ETH Zentrum, CH-8092 Zurich, Switzerland
phone: 4441632 5328, fax: 441632 10 86
E-mail: birgit.kopainsky@iaw.agrl.ethz.ch, christian .flury@iaw.agrl.ethz.ch,
peter.rieder@iaw.agrl.ethz.ch
Zurich, May 16, 2003
1 Introduction
Structural change in Swiss agriculture has been continuing over the last decade. This has
not only led to a decline in agricultural employment and income generation, but also to
land use changes with far-reaching consequences for the provision of landscape and ecol-
ogy related public goods. Most affected by this development are peripheral regions in the
Swiss Alps. They have at the same time experienced considerable support from regional
policy measures in the past. However, both agricultural and regional policy are undergo-
ing fundamental changes as a reaction to the effects of world trade liberalization, interna-
tional integration and increasing national budget limitations. The policy measures there-
fore become increasingly decentralized and restricted to innovative projects based on lo-
cal and regional initiatives. The task of local and regional decision makers has conse-
quently shifted from implementing national strategies to launching and moderating such
initiatives. A region's survival and quality of life no longer depend on the amount of
money paid by the Federation, but on the initiation and the appropriateness of autono-
mous development strategies. Hence, decision support tools are required both for evaluat-
ing the effectiveness and efficiency of new national or sectoral policy concepts and for
estimating the economic, social and ecological impacts and trade-offs of specific local or
regional development projects. Simulation and optimization models have proved useful in
providing decision support for the development of effective and efficient policy measures.
This paper explores the suitability of two different model types for addressing the above
mentioned issues. We distinguish between a linear programming model and a system
dynamics model with which we evaluate different development and policy scenarios for
agriculture in the Swiss Alps. The models provide insights into the processes underlying
agricultural development and their main influencing factors. We compare policy outcome
and policy contrasts by applying the two models to the same question, i.e. to the effects of
structural change in agriculture on enployment, income, land use and the provision of
public goods. We analyze which and whose questions can be asked to the models and
what answers and policy implications can be queried from the models, especially if the
questions are the same.
The questions arise from the characteristics and problems of Swiss mountain agriculture
and are illustrated in section 2. In section 3 and 4 we describe and compare the models
that were developed for the analysis of these problems. The results generated by the
models are displayed in section 5. Based on the comparison of model characteristics and
on optimization and simulation results, respectively, we draw conclusions concerning the
suitability of the two models for different decision makers and for applied research in sec-
tion 6.
2 Importance of agriculture in Swiss mountain areas
In the course of economic development, agriculture's productive contribution decreases
and the importance of the provision of public goods and of (positive) external effects in-
creases (Flury et al. 2002). In Switzerland agriculture contributed only one percent to the
gross value in 2000 using as much as 3.5 percent of the full-time workforce (SBV 2001).
Despite its declining economic role, agriculture still retains its spatial responsibility: more
than one third of the Swiss area is under agricultural cultivation (SBV 2001). Furthermore,
land use offers numerous examples of spatial environmental externalities (Bouman et al.
999). Soil erosion, loss of habitats, increased vulnerability of soils and loss of natural
amenities are manifestations of the negative effects of land exploitation. On the other
hand, appropriate land use activities prevent the emergence of these negative external-
ities. These types of spatial externalities are particularly pronounced in mountain areas
where, for instance, fallow land may generate negative external effects (Flury et al. 2000).
In these regions agricultural land use is also very important for landscape conservation as
it is a significant asset to the tourist industry. Additionally, agriculture has long been the
central economic sector in mountain regions, thus fostering and maintaining settlement,
economic and social life in remote areas (Errington 2000).
There is consequently considerable interest in the further development of agriculture, es-
pecially in mountain areas. Based on the above mentioned private and public goods pro-
vided by agriculture we were interested in the following three aspects: development of
land use, agriculture’s contribution to employment and to total value added.
We examined these aspects in a case study region in the Swiss Alps: The Muenstertal is
situated in the south eastern part of Switzerland. It consists of six villages situated be
tween 1200 and 1900 meters above sea level and is characterized by negative population
development and a high share of agricultural employment (23 percent on average). In the
year 2000 61farms were counted. Together, they cultivated a total of 1016 hectares. The
Muenstertal is delimited by the Italian border to the south, high mountains to the east
and west and the Ofenpass to the north, thus forming a well-marked regional system
with easily definable exchange relations to adjacent regions and the rest of the world.
3 Description and comparison of the models
3.1 Linear programming model
The linear programming model is a comparative-static sectoral optimization model (see
Figure ]) the objective of which is to maximize the sectoral revenue of agriculture in the
region under consideration. This is achieved by the optimal allocation of scarce production
factors to farm types in different sub-regions or communities, respectively, giving due
consideration to the restrictions of each level of aggregation (farm, community or region).
This means that the overall sectoral structure is optimized simultaneously in a way which
results in a maximum payment of production factors according to the criterion of
comparative cost advantage. A sector model helps to explain producers’ reactions to
external changes, such as prices or policy measures. Important information on the
producers’ expected reactions can be obtained by solving the model under different
assumptions on policy parameters or policy measures (Hazell and Norton 1986).
Figure 1 Structure of the linear programming model (RHS=Right Hand Side)*
Farm type 1
Farm type 2 R
Farm typei
Subregion 1
Farm type 1 H
Farm type 2
Farm typei
Subregion 2 Ss
Regional agricultural sector
Sectoral objective function
The main focus of the model is the detailed representation of land use. This allows for an
accurate ex-ante evaluation of different direct payment schemes. In addition, the model
includes indicators for the evaluation of different development and policy scenarios. These
indicators are some of the structural components of the linear programming model: ac-
tivities, restrictions and objective function (Figure 7). Sub-regions were defined conform-
ing to the borders of the six communities. Within each community, five farm types com-
pete for the available land. Grassland is used for different types of cattle (dairy or nursing
cows, beef or calf production, cattle rearing) and sheep. Various cropping activities are
possible (maize, barley, potatoes, wheat). Full-time and part-time farms are included and
land can be used at two different levels of intensity.
In the Right-hand side of the model spatial data are integrated as restrictions. The model
includes the following spatial data from different geographic information systems and
from public statistical databases on a per hectare grid:
¢ Agricultural land in each community, subdivided into six altitude levels (<600 m, 600-
3800 m in steps of 300 m, >1800 m above sea level) and four slope categories (<18%,
19-35%, 36-50%, >50%).
¢ Physical yields for each hectare and for different land use activities (grassland, hay,
silage, pasture), based on digital information of the cultivation suitability map and on
expert judgment.
¢ Input requirements (labor, mechanization, fertilizer) as a function of land use activity
(grassland, hay, silage, pasture) and of the topographic situation (altitude, slope).
Restrictions were mainly formulated on two levels of aggregation: 1) On the farm level all
productior-related restrictions are expressed (crop rotation, animal feeding, fertilization,
cattle and sheep rearing, labor etc.). Additionally, legal constraints for participation in di-
rect payments were included. 2) On the community level, restrictions concerning factor
markets were defined (available land on respective altitude level and slope category, bal-
ance of leased land and of labor).
In the optimization process for a given scenario, sectoral revenue (total of all farm reve-
nues) is maximized dmultaneously over all farm types and communities while at the
1For more details on the structure and components of the linear programming model see Appendix 1
same time adhering to the restrictions on the farm and community level. The total reve-
nueon the farm level is calculated as follows:
Revenue from agricultural production and direct payments
+ additional non-agricultural income
= direct costs (mineral fertiliser, feeding concentrates, seed, plant protection)
- machine and building costs
- costs for hired labor
revenue and costs of leased land
I+
= Total revenue on farm level
The model is programmed in AMPL (Fourer et al. 1993), an index-based programming lan-
guage. With the help of this software, the basic structure of the database (indices and
sets) is defined and the linear programming matrix is generated. The database is handled
independently from the model with an input generator in the relational database soft-
ware Microsoft Access 97 (Keusch 2000). The optimization is accomplished by applying
the CPLEX barrier algorithm (CPLEX User Guide 8.5.3) on a high performance parallel-scalar
computer server. Further details on the structure of the model and technical information
can be found in Flury (2002).
The original mode! was developed for the overall Swiss Alpine region. It was used as a
communication and synthesis tool within the scope of the multidisciplinary research pro-
gram “PRIMALP — Sustainable primary production in the Alpine region” at ETH Zurich (see
www.primalp.ethz.ch).
3.2 System dynamics model
The system dynamics model shows the agricultural part of a greater system dynamics
model on regional rural development. The purpose of the overall model is to test the €-
fects of economic development as well as of agricultural and regional policy scenarios on
the different functions of rural areas.’
The agricultural part represents livestock farming, land use and their combined effects on
and interrelations with employment and income generation. The model is conceived for a
time horizon of twenty years. This allows for the measurement of processes with a con-
siderable adaptation time like sectoral emigration of workforce and changeover in live-
stock farming (delays due to age of workforce and of stables, respectively). In Figure 2, the
subsystem diagram of the model is displayed, showing the network of development-
relevant issues faced by regional agriculture. The provision of public goods, of external
effects and agriculture’s contribution to settlement, economic and social life (see sec
tion 2) can be derived from the variables within the subsystems of land use, livestock and
agricultural workforce.
? See Kopainsky et al. 2003 for more information on the functions of rural areas.
Figure 2: Subsystem diagram of the system dynamics model?
AGRICULTURALPOLICY
* Direct payments
* Ecological constraints
¥
Lanpuse
+ Demandworkinghours — |¢ssesses
+ Supply fodder
+ Extensive& intensive use
* Production of public
‘goods (landscape &
biodiversity) ‘AGRICULTURAL
any bicoe seueep| EMPLOYMENT
He + Supply working
Livesrock ners
+ Demandworking hours
+ Demandfodder
+ Dairycattle & nursing
cattle
* Production of private es"
‘goods (milk & meat)
* +
PULLFACTORS
+ Regional
‘economy
SECTORAL
EMIGRATION
‘AGRICULTURAL POLICY DEVELOPMENT OF PRODUCT
* Direct payments PRICES
* Ecological constraints
The model distinguishes between two categories of livestock (dairy and nursing cows) and
three categories of land use (extensive pasture, extensive grassland and intensive grass-
land) some of which can fall fallow. The core part of the model involves a stock of agricul-
tural workforce which is drained by sectoral emigration. The model consists of four com-
ponents described below and was developed using the software “Vensim”.
¢ Livestock farming line: The two possible ways of livestock farming are shown: dairy
cows and nursing cows. The total number of cattle is delimited by ecological con-
straints, labor and fodder availability. Due to changes in national and international ag-
ricultural development, milk production becomes less and less competitive, especially
in mountain areas (Flury and Rieder 2003). Thus, the process only goes from dairy cows
to nursing cows and not backwards. The adaptation time of this process is determined
by the age structure of the stables in the region.
¢ Land use line: Land use changes are a direct consequence of changes in livestock farm-
ing (intensive to extensive grassland) and of the relation between total cultivated area
and total workforce (grassland to pasture with work becoming increasingly scarce).
Fallow land occurs when the maximum share of pasture is reached and work availabil-
ity continues to decrease.
3 For more details on the structure of the system dynamics model and its components see Appendix 2. The
Vensim model in stock and flow format is available from the corresponding author.
¢ Agricultural income: In this part of the model total agricultural income is calculated.
Important components of agricultural income are direct payments that remunerate
the provision of public goods.
¢ Agricultural employment: The stock of agricultural workforce is drained by sectoral
emigration. Emigration occurs at an adverse difference between the income generated
by livestock farming and land use and the potential non-agricultural income. The delay
in emigration depends on the age of the agricultural workforce and on the economic
performance of the overall regional economy. Another possible cause for emigration is
too big a discrepancy between demand and supply of working hours, i.e. when the ag-
ricultural workforce is not fully engaged with the available area and livestock in the
region. Again this process only works one-way, as agricultural enployment is highly
unlikely to increase given the high transaction costs of leaving agriculture and given
theory and evidence of the overall economic structural change and transformation.
3.3 Comparison of the models
In this section we compare the two models from a methodical point of view and evaluate
the consequences of the choice of a model type on the thematic questions raised in sec-
tion 2: development of 1) land use, 2) agriculture’s contribution to employment and 3) ag-
riculture’s contribution to total value added. As a basis for this Table 1reviews model
characteristics of both the linear programming and the system dynamics model.
Table 1 Comparison of thetwo models
Criterion for
Linear programming model
System dynamics model
comparison
Model purpose Synthesis of interdisciplinary research | Understanding of local rural deve-
on mountain agriculture, evaluation lopment processes, evaluation of re-
of agricultural policy measures gional policy measures
Focus on agricultural land use and Focus on stability and adaptation
sectoral development processes
1 Model type Comparative-static sectoral optimiza- | Dynamic regional simulation model
tion model
2. Model structure
Different farm types, communities
and overall region
One virtual farm for the entire region
under consideration
3. Time horizon
10 to 20 years
20 years
4. Underlying theory
and relevant
discipline
Microeconomic theory
Operations Research
Microeconomic theory
Systems Theory & System Dynamics
5. Determinants of
actor’s decision
Free decisions among possible input-
output-combinations (production
Decision rules
making functions) within restrictions
Decisions take into account price and Decisions are process-dependent, i.e.
cost prospects in the future decision at time t, depends on the
situation at timet,
6. Level of detail High due to model purpose Medium (higher level of aggregation)
Spatial disaggregation: spatial distri-
bution of agricultural production cap-
tured
due to mode! purpose
Spatial aggregation: capturing of
spatial distribution not possible due
to model structure (single virtual
farm)
Criterion for
comparison
Linear programming model
System dynamics model
7. Complexity and
scope of the model
Only quantitative variables
High number of variables, more the-
matic detail but less complexity due
to static approach
Interdisciplinary within agriculture
Quantitative and qualitative variables
Fewer variables but complexity due
to numerous feedback loops
Interdisciplinary in a broader sense:
agriculture as part of overall regional
system
8. Predictions
Prediction of optimal (from a sectoral
point of view) agricultural structures
that will develop after elapse of time
horizon of the model (after 10 to 20
years)
Measurement of scarcity of produc-
tion factors by means of shadow
prices
Critical factors for different policy
measures and production alterna-
tives
Prediction of structural adaptations
during time horizon of the model
(during 20 years)
Identification of irreversible pro-
cesses
Critical factors for system stability
9. Prediction preci-
Low to middle
Rather low, focus on system behavior
sion and not on single variables
10. Acceptance of the Positive in relation to acceptance and Positive in relation to acceptance and
model and credibil- | credibility: high level of detail and credibility: inclusion of qualitative
ity of the results high level of precision within as- variables, group model building and
sumptions about economic factors
Negative in relation to acceptance
and credibility: economic factors are
the only influences on behavior, “ex-
pert model”
communication tool, “common
model”
Negative in relation to acceptance
and credibility: low disciplinary preci-
sion due to holistic view
IL Transferability to
other regions
Little time intensive
Moretime intensive, as qualitative
variables have to be estimated sepa-
rately for every region
2 Work input for
model develop-
ment
Basically higher than system dynam-
ics model
Basically lower than linear program-
ming model
Dependent on share of qualitative
variables.
Criteria 1to 3 recapitulate model descriptions in the previous sections. Both models are
economic models (criteria 4 and 5). In the linear programming model, factors are allocated
according to neoclassical theory whereas in the system dynamics model, Johnson's theory
on quasi-fix factors (Johnson and Quance 1972) is taken into account. There are also differ-
ent implications of the economic feature of income-maximization. In the system dynam-
ics model, decision rules are formulated on this assumption. This implies sectoral emigra-
tion of agricultural workforce if the benchmark income is not achieved with agricultural
production or if there is an over-supply of working hours within agriculture. The system
dynamics model itself is a simulation model without optimization features. The linear
programming model, on the other hand, truly maximizes the sectoral revenue.
The level of detail covered by the models and the accuracy of the optimization or the
simulation predictions are one of the main influencing factors on model acceptance and
credibility (criteria 6 to 10). However, the more participatory and holistic procedure in the
system dynamic approach is probably of equal importance. Acceptance and credibility
thus depend on the target group of the model and their claims to simulation results.
Criteria Tand 22 refer to implications of mode! characteristics on mode! development and
model application.
In general, Table Limplies that the strength of the linear programming model is higher
prediction precision. The linear programming model, however, cannot capture adaptation
processes and over-estimates factor mobility, especially in the short term. It allows for the
optimization of structural reactions to a change in general conditions and for the estima-
tion of the consequences on income-generating production processes. The predictions are
not for a specific point in time. They much rather display the kind of structures that have
to be expected in the future given the simulated policy scenario and unrestricted factor
mobility. If the adaptation process is of crucial importance in Operations Research, recur-
sive-dynamic optimization has to be used. Such approaches and models were first devd-
oped by Day (1963) and De Haen (197). Recursive programming constrains the adaptation
process by including additional restrictions. For the optimization of structures at time t,
elements from the structure at time t, are integrated as flexibility coefficients. The reason
for this coupling is that actors’ decisions at a given point in time depend on their decisions
in the past (De Haen 197).
Covering adaptation processes, delays and irreversibilities is one of the main advantages
of the system dynamics model. System behavior, though, risks being rather sensitive to
the values of these system elements as general assumptions on factor flexibility have to
be made and decision rules have to be formulated.
4 Model validation
Once a linear programming model has been constructed, there are two broad ways in
which it may be in error: 1) the matrix may be inconsistent with the problem definition,
or 2) the problem definition may be incorrect or inappropriate in some way (Pannell 1997).
The linear programming model in this paper is considered as valid when it reproduces the
real agricultural structures in the year 2000 in a sufficiently accurate way. Table 2 com-
pares model results for the year 2000 (calibration) to the real situation in this year. The
data used for the validation are based on information on the situation of the year 2000,
obtained from different statistical databases on Swiss agriculture and from expert know-
ledge about the case study region.
Table 2: Validation results of the linear programming model
Variable Real Calibration
Agricultural workforce (persons) 532 4B
Total area (hectares) DV DV
Extensive land (hectares) 227 19
Units dairy cattle 389 364
Units nursing cattle 841 626
Agricultural income per person (CHF) 68363.2
The comparison of the real structures and the calibration results shows that the model
provides a reasonable reflection of the structural reality. All the agricultural land is culti-
vated. This is in keeping with the actual situation where no fallow land exists. The low
number of nursing cattle can be explained by the observation that nitrogen and phospho-
rus nutrient balances are more binding in the model than in reality. Information on the
actual agricultural income is not available on a regional basis and can only be estimated. A
value for this variable is therefore not provided for the real situation.
The validation process of the system dynamics model included testing model structure as
well as model behavior (Barlas 1996). Structure validation is concerned with warranting
that the model's internal structure is a sufficiently accurate description of the real system,
with respect to the issue of interest. Behavior validation means that the output behavior
of the mode! reproduces closely enough the dynamic behavior of the real system under
study (Barlas et al. 2000). For this purpose the accurate replication of the patterns for the
period between 1996 and 2000 was tested (same data source as for the validation of the
linear programming model). This period is rather short. There are, however, problems with
data availability for the time before 1996. Additionally, Swiss agricultural policy has ex-
perienced profound changes since 1992 with far-reaching consequences on agricultural
activities and actors’ decision making. It is therefore difficult to compare the situation be-
fore the agricultural policy reform with the current situation without completely altering
the processes covered in the model.
Tests on model structure showed high sensitivities to the values of the adaptation times.
Exogenous variables proved little sensitive. Assumptions on endogenous relations like
change in type of livestock and especially change in type of land use, on the contrary, had
considerable impact on system behavior. This is well in line with the system dynamics no-
tion of causally closed systems where the internal structure is much stronger to explain
behavior than external influences (Richardson 1991).
Results from behavior validation are shown in Table 3. Although the absolute values of the
variables differ between the simulation results for the period between 1996 and 2000 and
the observed development, all the processes except nursing cattle go in the same direc-
tion and show identical behavior. The extreme difference for nursing cattle can be e-
plained by the fact that total cattle intensity in 1996 is much higher than ecological con-
straints from agricultural policy would allow for and that the model only permits a correc-
tion via nursing cattle. The reaction cannot be balanced within the short simulation time
of four years.
Table 3: Validation results of the system dynamics model
Variable 1996-2000 (model) 1996-2000 (reality)
Agricultural workforce -28% -3%
Dairy cattle -20% -60%
Nursing cattle -90% +290%
Extensive pastures (+155 hectares) (+1Lhectares)
Fallow land 0 ie)
5 Model results
In this section we present optimization results from the linear programming model and
dynamic patterns which the system dynamics model generated. We distinguish between
the evaluation of effects of different policy assumptions for direct payments as a sectoral
policy measure and the evaluation of an endogenous devdopment strategy based on re-
gional marketing and agricultural self-help. We analyze model results with the aim of
identifying further differences in the two modeling approaches and in policy implications.
Therefore, we do not comment on theresults in every detail.
5.1 Model parameters and scenarios
Wesimulated and compared four scenarios:
1 Local processing: In the face of decreasing product prices the strategy of local
processing and marketing has gained in relevance. This holds especially for mountain
areas where consumers not only value product quality, but also the natural and socio-
cultural assets that the region’s name implies. The local processing scenario
investigates the effects of a local development initiative where higher product prices
can be realised through local processing and marketing of milk and meat.
2. New direct payments: All farmers in the mountain areas receive base payments of CHF
1200.- per hectare. A main motivation for these payments is to ensure farmers’
incomes in view of decreasing commodity prices. With the re-examination of Swiss
agricultural policy a new type of direct payments is discussed, especially for mountain
areas. Direct payments coupled to agricultural workforce instead of agricultural land
aim at stabilising agricultural employment and thus at fostering agriculture’s
contribution to settlement, economic and social life. In the new direct payments
scenario we study the effects of a change in the direct payments scheme.
3. Combination: For this scenario we combine the effects of the local processing and the
new direct payments scenario.
4. The results of the previous scenarios are always compared to base run where current
development trends remain unaltered in the future.
The data used for the simulations are the same as those used for model validation (sec-
tion 4). The essential exogenous parameters used in the two models are given in Table 4.
The assumptions for the future values of the parameters are based on estimations of the
Swiss Federal Agency for Agriculture.
Table 4: Assumptions and parameters for the different scenarios
Scenario Assumptions and parameters
Base run Cost development (index for costs land use and costs livestock):
= 1in 2000, 105in 2010, 11in 2020
Price development (index for milk price and meat price):
— 1in 2000, 0.9 in 2000, 0.8 in 2020
Development of potential non-agricultural income (index):
— 1in 2000, 105in 2090, 11in 2020
Direct payments: constant
Local processing Differences to base run:
— Price index milk: index base run*1065
— Price index meat: index base run*105
New direct payments Differences to base run:
— Decrease of base payments area from CHF 1200.-/ha to CHF 600.-/ha
between 2008 and 202
— Introduction of direct payments workforce starting with CHF 3'000.-
/person to CHF 12'000.-/person between 2008 and 202
Combination Combination of assumptions for local processing and new direct payment
5.2 Simulation and optimization results and thematic comparison
In Figure 3 the results from the linear programming model for the four scenarios are dis-
played. The results are given as percentages of the calibrated situation in the year 2000
(see validation). Figure 4 shows the dynamic patterns generated by the system dynamics
model. Again, the results are percentages of the real situation in the year 2000 (year 0 of
the simulation). As the simulations in the system dynamics model produced virtually iden-
tical outcomes for all four scenarios only base run results are displayed. The only variable
that differed between the scenarios, agricultural income per person, is depicted in Figure
5.
Figure 3: Optimization results from the linear programming model for the four scenarios
0%
160%
40%
Base run
120%
100% New direct
payments
80% 4
Local processing|
60%
40% [I Combination
Percentage of Calibration (=100%)
20%
0%
Agricultural Totalarea Extensive Unitsdairy Units Agricultural
workforce (ha) land (ha) cattle nursing income per
(persons) cattle person (CHF)
Figure 4: Simulation results (base run) from the system dynamics model
20%
~e Agricultural
100% Workforce
a (persons)
80% i+ Units dairy
ia cattle
60%
t= Units nursing
cattle
40%
— >< Extensive
20% pastures (ha)
0%
Percentage of Year 0 (=100%)
Time (Year)
A comparison between the results of the linear programming model and the situation for
year 20 in the system dynamics model shows a high degree of congruence. Agricultural
workforce is reduced to a bit more than 60 percent and livestock experiences a marked
shift from dairy to nursing cattle. Fallow land arises to a minor extent (not shown in Fig-
ure 4). A difference lies in the development of agricultural income per person. Whereas it
remains more or less constant in the linear programming model it increases steadily in
the system dynamics model (see Figure 5). The relative differences in income per person
for the four scenarios are similar for both models. The difference in the development of
the variable between the two models arises as overall agricultural income declines more
than agricultural workforce in the linear programming model. The opposite holds for the
system dynamics model. Additionally, due to model structure and the limited level of de-
tail,incomein the system dynamics model is rather high in general.
Because the system dynamics model generates almost identical behavior for the four sce-
narios a general feature of complex systems is confirmed, namely that they resist most
policy changes (Forrester 1971in Richardson 1991, Sterman 2000).
The linear programming model, on the contrary, shows different and also oppositional
results for the four scenarios. New direct payments as well as local processing result ina
higher share of dairy cattle than in base run. Dairy cattle even increase to 18% of their ini-
tial value when new direct payments and local processing are combined. This is a direct
consequence of the income-maximization feature underlying the model: As product
prices for milk increase (they increase slightly more than for meat, see Table 4) and as
work is subsidized by the new direct payments, the work intensive cattle type (dairy cattle)
is chosen. With a decline in land-related payments work extensive areas are abandoned
which leads to fallow land of almost 10%.
These results cannot be achieved in the system dynamics model for two reasons. First the
system dynamics model is a simulation and not an optimization model. Second, in the
new direct payments scenario the change in the payment scheme occurs from year 8 to 2.
By this time, however, some persons have already left agriculture and the change in live-
stock from dairy to nursing cattle has well proceeded. We find here an example of path-
dependency. Those persons that have left agriculture will not return as they have under-
2
gone re-education, found a job elsewhere in the regional economy or emigrated from the
region. The same holds for livestock. Once the stables for dairy cattle are converted into
stables for nursing cattle, this process will not be reversed. The desired effects of the new
direct payments scheme can therefore not set in. This is a crucial result from model com-
parisons and will be discussed further in the conclusions.
Figure 5: Income development in the system dynamics model for the four scenarios
20000
110000
100000
90000
80000
Agricultural income per person
70000
60000
o 12 345 678 9D URBH EB KB VY B BP 2
Time (Year)
baserun ~~ - -local processing © — - - new direct payments © — — combination
6 Conclusions
In this paper we compared policy and outcome contrasts of a linear programming model
and a system dynamics model. The models were both applied to the question of the e-
fects of structural change in Swiss mountain agriculture on regional land use, agricultural
employment and agricultural income. We studied different policy measures in order to
analyze model behavior and policy implications of model results. We thus aimed at identi-
fying and differentiating the suitability of the two model t ypes for decision support.
From the methodical comparison of the two models conclusions can be drawn as to which
and whose questions can be asked to the models:
¢ The linear programming model generates information on the development of the ag-
ricultural sector as a reaction to changes in economic, social or political conditions. It is
therefore especially suited as a decision support tool for policy makers elaborating na-
tional policy concepts and optimal sectoral policy measures for given goals.
¢ The system dynamics model is most useful in a specific region. It is valuable in regional
and endogenous development initiatives as a communication tool and provides local
and regional decision makers with holistic understanding of the crucial factors affect-
ing regional rural development in the long run. It helps deciding on the effectiveness
of additional local and regional measures for specific development strategies given the
overall sectoral development. A system dynamics model is also suited as a decision
support tool for the regionally adapted implementation of national policy concepts
and of sectoral policy measures.
In addition to the methodical comparison, the comparison of model results for the differ-
ent policy scenarios allows conclusions as to what answers can be queried from the mod-
els if they are applied to the same questions. The analysis of the effects of alternative pol-
icy scenarios showed that different results and thus different policy implications are gen-
erated if the policy measure includes a structural break. This could be observed with the
new direct payment scheme (combined with local processing) that was tested for the goal
of maintaining dairy farming and thus agricultural employment in mountain areas. In this
case, the policy recommendation drawn from the linear programming would be to subsi-
dize agricultural workforce. The system dynamics model identified the change in livestock
from dairy to nursing cattle as a path-dependent process triggered mainly by product
price signals. The policy recommendation would consequently be to stabilize dairy farm-
ing via milk price.
It is therefore advisable to combine the two model types for decision support in applied
research. The linear programming model specifies optimal future structures for a given
policy goal. Whether these can at all be achieved is answered by the system dynamics
model. It identifies the path-dependent processes that influence the effectiveness and
efficiency of optimal policy measures and provides the information necessary for a suc-
cessful communication andimplementation of the measures.
Even if the models are not combined in practical application, both modeling aoproaches
can profit from each other. The benefit of a linear programming model for the elaboration
of a system dynamics model is accurate information on the overall development chal-
lenges faced by the regional system. These challenges are subsequently addressed by the
system dynamics model so that the constraint of restricted precision can be considerably
alleviated. The benefits of a system dynamics model for a linear programming model, on
the other hand, lie most of all in the coverage of the adaptation processes and in the iden-
tification of path-dependencies. Another important benefit is that the purely quantitative
economic model can be expanded by the qualitative variables included in the system dy-
namics model.
Acknowledgements
We are grateful to Kurt Zgraggen for his valuable suggestions and to three anonymous
reviewers for helpful comments.
References
Barlas Y. 1996. Formal aspects of model validity and validation in system dynamics. System
Dynamics Review 2 (3): 183-210.
Barlas Y., Cirak K., Duman E 1996. Dynamic simulation for strategic insurance manage-
ment. System Dynamics Review 56 (1): 43-58.
Bouman B.A. M., Jansen H. G. P., Schipper R.A., NieuwenhuyseA., Hengsdijk H., Bouma].,
1999. A framework for integrated biophysical and economic land use analysis at dif-
ferent scales. Agriculture, Ecosystems and Environment 75: (1) 53-73.
Day RH. 1963. Recursive programming and production response. North-Holland Publish-
ing Co., Amsterdam.
De Haen H. 1971 Dynamisches Regionalmodell der Produktion und Investition in der
Landwirtschaft. Agrarwirtschaft Sonderheft 43. Alfred Strothe Verlag, Hannover.
Errington A. J. 2000. Rural development and the rural economist. In: Hillebrand H., Goetge-
luk R, Hetsen H. (eds.): Plurality and rurality. The role of the countryside in urbanised
regions. Agricultural Economics Research Institute (LEI), The Hague. Vol. 2/2: 115-128.
Fourer R., Gay D. M., Kernighan B. W. 1993. AMPL A modeling language for mathematical
programming. Body & Fraser, Danvers MA.
Flury C., Gotsch N., Rieder P 2000. Socio-economic and ecological effects of alternative
direct payment regimes on different Swiss Alpine regions. Cahiers d'Economie et So-
ciologie Rurales 57: 5-26.
Flury C. 2002. Zukunftsfahige Landwirtschaft im Alpenraum - Entwicklung von Nutzungs-
strategien fiir den Kanton GraubUnden auf der Basis eines Sektormodells. Wissen-
schaftsverlag Vauk, Kiel.
Flury C., Rieder P. 2003. Strukturelle Auswirkungen auf das Berggebiet, Folgestudie Umla-
gerung der Milchpreissttzung, Teil 3, Institut fUr Agrarwirtschaft, ETH Zurich.
Hazell P. B.R., Norton R. D. 19986. Mathematical programming for economic analysis in ag-
riculture. MacMillan Publishing Company, New York.
Johnson G. L, Quance C. L 1972. The overproduction trap in U.S. agriculture. A study of re-
source allocation from World War | to the late 1960's. Resources for the Future, Inc.,
Washington D.C.
Keusch A. 2000. Modellierung ressourcendkonomischer Fragestellungen am Beispiel der
Erosion im Gebiet des Baldeggersees. PhD dissertation, ETH Zurich.
Kopainsky B., Buchli S., Rieder P. 2003. How peripheral and agrarian communities work and
how their viability can be improved - insights from a System Dynamics approach.
Paper presented at the Regional Studies Association International Conference ,,Rein-
venting Regions in a Global Economy“, Pisa, April 12-5, 2003. http://www.regional-
studies-assoc.ac.uk/ events/ pisa0 3/ kopainsky.pdf
Richardson G. P. 1991 Feedback thought in social science and systems theory. University of
Pennsylvania Press, Philadelphia.
Pannell D. 1997. Practical linear programming.John Wiley &Sons, New York.
SBV (Schweizerischer Bauernverband) 2001 Statistische Erhebungen und Schatzungen
Uber Landwirtschaft und Ernahrung. Brugg.
Sterman J. D. 2000. Business Dynamics. Systems thinking and modeling for a complex
world. McGraw-Hill, Boston.
Appendix
Appendix 1 Detailed overview of the structure of the linear programming model
RHS RHS sub-
Activity farm level} | regional level
g
g n
5
ra om Bed
oO o|s
2/5 (E/218
S12 815 je 5
ale [Slo |S ule 8
£l/S5(2/S |e wis pL}
fom GIF i slo lo ©
a/9 BIlE/8I> (8/8 «| (El) al/8
E/SILIElolLZisle oO elalols
E Gie fe fol is 5|o Sloa|al=
2ZIis/z S18 /S\i5/s ls £/a\o S)£/2/5
Level of BP SIZ IZISIsiZisl /sleis] (slslsls
couree- Slo /o/8/8 [8/5/82] le [>| 8| |e 2 [Els
1ggrega , Z/SlSISIEISIZIS IS] IZ VEIE| SSElEIS
ton Constraint 612 |S la la |e |e |e |5 | |S 18 || 15 lA LEIS
Land use x |x x x
Crop rotation x
Animal feeding xe [oe x
Farm Fertilization x |x] x x
Cattle and sheep rear- x x
ing
Milk production x x x
Labor x| x} x] x]x x x
Market for leased land x
Commu- _ | Labor market x x
nity Land use BEX x
Summering pasture x x
Sector [Animal trade TT TTT rt tt i tt et 7
Appendix 2: Detailed overview of the structure of the system dynamics model (simplified
stock flow diagram)
POTentialincome ror
non-agricultural
employment per person
Tscrepancy income
(agr vs non-agr)
-dncome from
agrictturliffome land use>
grea
Workforce Full perperson agheulturaf#~
a Employed [we income
MIGRATION emigration, > __- FS ctncometrom
DELAY livestock
<DiRect payments
word HouRsPER wworforce>
FULLY EMPLOYED
aren -erorkng hours total cultivated
working hours (5) Nppiywotking Jenaiser oe
hours ge. svorkinghours
demand wore ———_lvestock> oat Nel per total
ours workforce
effect of area per
person on change in Bo cndonment
DELAY IN CHAN FaoaMtuM AREA pret
OF LIVESTOCK EXTENSIVE GRASSLAND nN
HOURS PER UNIT PER PERSON, “ADAPTATION TIME
> PERSON WHEN DAIRY LaNDUSE
DAIRY CATTLE — scat to milf income PARSING effect of area per person
hea is ‘onchangein extensive
-priceindex Tanduse gassend to
milk pasture extensive
EASE PAYMENTS “Sncombper uit nursing cat
Cartes ee imcomeperunit | HodRs PER UN? FODDER DEMAND PER
dairycattie | yyRsinG Carmi UNIT DAIRY CATTLE
coms dal apaptaniod ME
‘cattle <Mllkproceeds> income per unit NURSING CHANGE MAXIMUM CA
ek AT ae apeteatston fODnes Dena Pe
ae clndindbe cadleitensityon GRASSLAND “
v
Back to the Top
CC BY-NC-SA 4.0