Online Supplementto “Land-use change as the result of rural land exchange: an agent-based simulation model”
This supplement contains background information on the following topics:
- Observed land use in 2001 and 2009
- Farmer agent behaviour
- Parcel evaluation by farmer agents
- Parcel evaluation by nature agents
- Parcel evaluation by estate owner agents
- Minimum transaction price
- The area a farmer agent can buy
- Intensifier/innovator growth
- Trends in economic size per farming type
- Technical details of RULEX
- Maps of simulated land use in 2009 and an overlay between observed and predicted land use change
Analyses and modelling were carried out with the following software packages:
-RULEX is developed in Java and uses the Repast Simphony 2.0 ( suite;
-Statistical analyses were done using R 2.13.8 (
-Spatial operations were done in ArcGIS v.10.2.
1. Observed land use maps in 2001 and 2009
By linking census data for farms to the spatial parcel database, we could assign farming type to each parcel. We combined this with spatial data of properties of nature organizations and estate owners. That data was only available for 2009, so for setting up RULEX, the spatial delineation of properties of estate owners and nature managers for 2001 was obtained by subtracting the parcels that belonged to farmers in 2001. Here we show only the 2001 data that was available, which concerns the farming type for those parcels of which the owner was recorded in the census of 2001. For 2009 we show the combined data of the farming type and the nature organization and estate owner properties.
Figure s1. Farming type in 2001 (top) and land use (combined farming type and nature-organization and estate-owner properties) in 2009 (bottom). The ‘nodata’ category in the 2009 situation comprises built-up land and infrastructure; in the 2001 situation the nodata category contains nature areas, estates, and unregistered agricultural parcels as well. Nature areas and estates in 2001 were inferred by taking the nature areas and estates of 2009 and subtract the areas that were registered in 2001 as belonging to a farmer.
2. Farmer agent behaviour
The calibration of the function that describes the behaviour probability conditional on farm(er) properties was done as follows: Using the farm panel data, we classified farmers into the four behavioural categories (expanding, selling, intensifying/innovating, or stable). The agriculture census farm data of 1999 and 2009 formed this panel database. Not all farms present in 1999 could be linked to the farms present in 2009 and vice versa, so that the panel database contained 740 records.Farms that purchased more than four hectares of landin the period 1999-2009 were classified as expanders, farms that decreased in both economic and physical size by more than 20% were classified as shrinkers, farms that grew in economic size and also increased intensity (economic sizeper hectare) while not being classified as expanders were classified as intensifiers, and the remaining farms were classified as stable. In order to arrive at these criteria, we explored a wider range of threshold values, and these turned out to give best results in terms of (a) maintaining a sufficient number of observations in each category and (b) the best score in the below-described regression analysis, indicating that groups were most distinguishable using that set of criteria. These criteria resulted in the following distribution of farmers in the panel dataset: 38% expanders, 18% shrinkers, 8% intensifiers, and 37% stable farms.
Next, we performed multinomial logistic regression analysis to calibrate functions that describe the probability of showing these behaviours, conditional on the farmer’s age, physical farm size, economic farm size, and farming type, all in 1999, and having been an expander (yes/no) in the preceding period (1989-1999). We used these variables, as well as log-transformations of size variables (which showed a strongly skewed distribution) and interaction terms of size variables and farming type. We applied a backward selection method to remove variables for which the coefficient could not be reliably assessed (p > 0.05).
Note that the probability functions in Eq. 1 in the main article include the variable farming type (more specifically, dairy farming versus the rest), but only in interaction with economic size (NGE). Being a dairy farmer or not does not affect the behaviour probabilitydirectly, but it does influence how economic size (NGE) affects the behaviour probability. For all non-dairy farmers, having a higher NGE results in a higher probability of being an expander or an intensifier, but for the dairy farmers this is slightly different: the higher the NGE the more they are inclined to expand (even more than other farmers) while a higher NGE will not increase their probability to intensify. This relates to manure legislation in the Netherlands: there is a legal maximum in the amount of manure that may be applied per hectare, so that farmers need to have a minimum amount of land per animal in order to be able to spread the manure over their own land. Because most farmers already are at this maximum, increasing the number of animals without increasing the area of land (i.e., intensifying) is usually not possible. For this reason, dairy farmers are, more than other farmers, dependent on being able to expand.
The computed McFadden’s pseudo R2 for these models was 0.16. Because there are only four categories, about 25% is correctly simulated due to pure chance. Hence, the regression model reproduced 41% of observed behaviours in the calibration dataset correctly. The distinction between shrinkers and stable farms was difficult as well as that between intensifiers/innovators and expanders. When we grouped these behaviours, the model reproduced 62% of the lumped behaviours correctly. For the RULEX model application this implies that in the course of a simulation some farmer agents tend to alternate between expanding and intensifying/innovating behaviour while others alternate between stable and shrinking behaviour.
3. Parcel evaluation by farmer agents
The price a farmer is willing to pay for a parcel (WTPF) is assumed to decline linearly with the distance between a parcel and the farmer’s farmstead. Estimating the coefficients of such a function from empirical transaction data is complicated by the fact that only the successful negotiations are revealed: high prices paid for nearby parcels are frequently observed while low prices paid for remote parcels are mostly absent. Therefore, we performed a Monte Carlo simulation of a virtual land market, and we calibrated the parameters that control this land market by fitting the simulated joint distribution of prices and distance to an observed distribution. Hereto we used the simulated annealing (‘SANN’) option in the ‘optim’ function in R (R Development Core Team, 2006). More details will appear in a forthcoming publication. The observed distribution was obtained from the records of rural land exchange, whereby we excluded all parcels that had structures on them or that were under long term lease contracts. This way, the slope of this line was estimated at -8, suggesting that with every 100 m extra distance, the WTP/Adrops by 800 euro ha-1. This gives rise to very local land markets, as also observed by Cotteleer et al. (2008).
Apart from this distance effect, parcels also vary in value due to parcel properties. In order to estimate these effects, we undertook a multiple regression with observed transaction prices as the dependent variable and the following independent variables: parcel size, parcel shape (perimeter to area ratio), soil quality, distance to major roads, distance to villages and towns, being situated inside or outside the areas designated for the National Ecological Network, and the distance to this network. For all distance measures, we tested both a linear relationship as well as an inverse-distance relationship. We controlled for the effect of ‘distance to farmstead’ by taking a subset of transactions in which parcel and buyer were less than 500m apart.This subset contained 228 observations.
We fit a gamma distribution as the parcel value has a tail towards higher prices. Table s1 shows the parcel properties that remained in the equation after backward elimination of non-significant parcel properties.
Table s1. Parcel properties influencing the parcel price
Parcel property / Coefficient estimate / p-value(intercept) / 4.16E-05 / 4.29E-07
Soil Suitability / -1.89E-07 / 4.60E-02
Inverse distance to towns and villages / -1.62E-03 / 3.42E-12
Inverse distance to areas designated for the National Ecological network / 5.26E-04 / 5.32E-03
The McFadden pseudo R2 of this regression is 0.25.
The obtained coefficients can be combined with the distance coefficient to result in Equation 2a (copied below):
WTPF = (4.16E-05 - 1.89E-07 * SoilSuit - 1.62E-03 * (1 / (dist_town + 100)) + 5.26E-04 * (1 / (nen_dist + 100)))-1–16.85 * (dist_farmstead - 250) (€ha-1)
Where dist_farmstead is the distance between the parcel and the farmstead (m), SoilSuit is an indicator for the percentage of the potential yield that can be obtained on the soil in percentage point, dist_town is the distance from the parcel to the nearest town (m), and dist_nen is the distance between the parcel and the National Ecological Network (m). Note that the same function is used to compute the WTAF for selling farmers.
The interpretation of the coefficients is not straightforward as, besides the transformations we made to the variables, the gamma link function introduces interdependence of variables and further non-linearity. Applying this equation to actual transaction records in the area indicated that: a) a decrease in one unit of soil suitability reduces the WTPF by approximately 300 euro ha-1; and b) the proximity to the national ecological network (NEN) also suppresses the WTPF (the difference between a parcel bordering the NEN and a parcel 100 m away from the NEN could amount to 3200 euro ha-1, while this effect levels off with larger distances). Distance from towns and villages had a strong positive effect on WTPF, due to farmers speculating on a change in the designation of the parcels for housing development. A parcel bordering the built-up area could be three times more expensive than a parcel that is 1 km further away, which strongly exceeds normal agricultural market prices.
4. Parcel evaluation by nature agents
Aquestionnaire based on the conjoint analysis approach was used to derive the WTP fornature agents. We asked six representatives of two different nature organizations (12 in total), who were all involved in parcel evaluation before acquisition, to order five fictional parcels from high (1) to low (5) preference.These five parcels differed in soil type, groundwater level, presence of exfiltrating groundwater, and distance to an existing nature area. An example of the representation of these fictional parcels is shown in Figure s2. All parcels represented a range of realistic values for the study area, in various combinations. Next, a regression was performed with preference (1 to 5) as dependent variable, and soil type (four categories), mean spring groundwater level (cm from surface), presence of exfiltrating groundwater (0 or 1), and distance to an existing nature area (m) as independent variables. This resulted in the following regression equation:
Preference score = 1.84 + 1.13E-3*distance - 1.32*seepage + 4.58E-2*groundwaterlevel
In order to rescale this equation to WTPN, the following steps were taken:
-The obtained preference score was subtracted from 5, in order to arrive at a measure that has low values for non-preferred parcels;
-All factors are multiplied by a scaling factor of 6980, in order to arrive at euro ha-1. This number was chosen in such a way that the range of paid priced generated by the resulting equation agrees with national statistics of average prices paid by nature managers for agricultural parcels.
This resulted in the following equation for WTPN:
WTPN = 22057 - 7.9*distance + 9214*seepage - 320*groundwaterlevel(€ ha-1)
This equation indicates a negative influence of distance to existing property. For every 100m further away from an existing nature reserve, the WTPN drops by 790 € ha-1. Within the range of distances we tested in the conjoint analysis (up to 2 km), we found no indication that this amount levels off. The presence of exfiltrating groundwater on a parcel was found to increase the WTPN by 9214 € ha-1. With every 10 cm increase in groundwater level, the WTPN increases with 3200 € ha-1. Within the ranges of groundwater level we tested in the conjoint analysis (between 45-100 cm), which correspond to the upper range of observed groundwater levels in the area, we found no indication of this price effect levelling off.
Because the scaling from preferences to euro ha-1 was not based on data specific to the research area, we derived the intercept of this equation from fine-tuning as discussed in section 4.1 of the main article.
Figure s2. An example of one of the maps with five fictional parcels that were ranked by representatives of nature organizations. Colour shades from blue to green indicate high to low groundwater levels; dark green polygons indicate an existing nature organizationproperty; the pattern of the parcel indicates soil type (following the Dutch national soil classification system); asterisks indicate whether there isseepage of exfiltrating groundwater or not.
5. Parcel evaluation by estate owner agents
In interviews, estate owners indicated that they would only buy land directly adjacent to their estate, regardless of physical characteristics. It was not possible to identify estate owners as buyers from the transaction database and so, this statement could not be verified from independent data. Hence, broad rules were implemented based on whether or not available parcels share a boundary with an existing property holding. Based on a few estate owners that could be identified in the transaction database, these agents were assumed to want to purchase adjacent parcels with a WTPE set at 35,000 € ha-1.
6. Minimum transaction price
Based on the observed transaction data, we found that about 99% of all parcels were sold at prices higher than 16,914 euro ha-1. Based on this observation, we assume that for a transaction to take place the minimum offer a shrinker receives from an expanding agent must be 17,000 euro ha-1. We assume the remaining 1% of transactions below this threshold concern transactions between relatives; something RULEX does not incorporate.
7. The area a farmer agent can buy
Over the period 1999-2009, all farmers classified in the panel database as expanders show a frequency distribution of expansion rate as shown in Figure s3.
Figure s3. Frequency distribution of expansion rates of expanders
Hence, the vast majority expanded by less than 100% over a period of 10 years. A few expanded more, but expansion rates of more than 250% were hardly ever observed. Based on this, we estimated that annual expansion rates do not exceed 30% of the physical farm size at that moment.
8. Intensifier/innovator growth
Farmer agents that intensify or innovate benefit from an increase in their economic size of 0.19 NGE per year. This value was obtained by taking the average annual increase in economic size of all farmers in the panel database that were classified as intensifier/innovator.
9. Trends in economic size per farming type
Trends in economic size (NGE) per farming type were estimated from price developments for important agricultural products[1]. Price changes approximate to changes in gross margins, which are in turn linearly related to our definition of economic size (in NGE). Trend lines were fit to price developments between 2001 and 2009 for milk, wheat, barley, potato, sugar beet, pigs for slaughter, flower bulbs and tree nurseries. We expressed these trend coefficients as percentages of the average price. The annual NGE trend for arable farming was taken to be the average of the trends for wheat, barley, potato, and sugar beet; for dairying it was taken to be the trend in milk prices; for pig-breeding it was taken to be the trend in ‘pigs for slaughter’ price; for horticulture it was taken to be the average of the trends for flower bulbs and tree nurseries; and for mixed farming we took an average of all four sector trends. In doing so, we arrived at the following annual NGE trends: Dairy-1.0%; Arable +1.4%; Horticulture -0.4%; Pig-breeding 0.7%; and Mixed 0.2%.
10. Detailed information of RULEX
In this section we present more detailed information about RULEX in the form of a pseudo code (Figure s4), a UML activity diagram (Figure s5) and a list of RULEX parameters and their default values (Table s2).
Pseudocode: RULEX main simulation scheduleInitialization:
1. Load model input data (see Table 1) for farmers, nature managers and
land parcels
2. FOR each agent ∈ Agents //all farmer agents and nature manager agents
i. Fill-in missing attributes, if any (see appendix)
ii. IF agent is Farmer THEN
call update-NGE(current-time, agent)
3. FOR each parcel ∈ Parcels
call update-biophysical-attributes(current-time, parcel)
Main Schedule: Runs for n time steps (years).
1. FOR each farmer-agent ∈ Agents
call update-strategy(current-time)
2. FOR each farmer-agent ∈ Agents
i. call update-strategy(current-time)
ii. call update-NGE (farmer)
iii. FOR each parcel owned by farmer-agent
call calculate-perceived-value(farmer-agent, parcel)
3. FOR each farmer-agent ∈ Shrinkers
i. rank all owned parcels based on perceived-value
ii. FOR each parcels tagged as least-desirable
call put-parcel-for-sale(farmer-agent, parcel)
4. FOR each agent ∈ Expanders NatureManagers
i. FOR each parcel available for sale
call calculate-perceived-value(agent, parcel)
ii. rank all parcels available for sale based on perceived-value
iii. FOR each parcel tagged as desirable
call place-bid (agent, parcel)
5. FOR each parcel available for sale
i. winner call choose-winner(parcel, parcel.bidders)
ii. call transfer-ownership(parcel, winner)
6. FOR each agent ∈ Agents
IF agent is Farmer THEN
i. call update-state(current-time, farmer-agent)
ii. IF farmer-agent is Dead or Retired THEN
call succession(current-time, farmer-agent)
IF agent is NatureManager THEN
call update-budget(current-time, agent)
7. FOR each parcel ∈ Parcels
call update-biophysical-attributes(current-time, parcel)
END
Figure s4. RULEX pseudo code