|Home | Archives | About | Login | Submissions | Subscribe | Contact | Search|
Copyright © 2011 by the author(s). Published here under license by The Resilience Alliance.
Go to the pdf version of this article The following is the established format for referencing this article:
Boone, R. B., K. A. Galvin, S. B. BurnSilver, P. K. Thornton, D. S. Ojima, and J. R. Jawson. 2011. Using coupled simulation models to link pastoral decision making and ecosystem services. Ecology and Society 16(2): 6. [online] URL: http://www.ecologyandsociety.org/vol16/iss2/art6/
Research Using Coupled Simulation Models to Link Pastoral Decision Making and Ecosystem Services
1Natural Resource Ecology Laboratory and the Department of Forestry, Rangeland, and Watershed Stewardship, Colorado State University, 2Department of Anthropology and the Natural Resource Ecology Laboratory, Colorado State University, 3Agricultural and Forestry Experiment Station, University of Alaska, 4International Livestock Research Institute, 5University of Edinburgh, 6Department of Forestry, Rangeland, and Watershed Stewardship, Colorado State University
Historically, pastoral people were able to more freely use the services their semi-arid and arid ecosystems provide, and they adapted to changes in ways that improved their well-being. More recently, their ability to adapt has been constrained due to changes from within and from outside their communities. To compare possible responses by pastoral communities, we modeled ecosystem services and tied those services to decisions that people make at the household level. We created an agent-based household model called DECUMA, joined that model with the ecosystem model SAVANNA, and applied the linked models to southeastern Kajiado District, Kenya. The structure of the new agent-based model and linkages between the models are described, and then we demonstrate the model results using a scenario that shows changes in Maasai well-being in response to drought. We then explore two additional but related scenarios, quantifying household well-being if access to a grazing reserve is lost and if access is lost but those most affected are compensated. In the second scenario, households in group ranches abutting the grazing reserve that lost access had large declines in livestock populations, less food energy from animal sources, increased livestock sales and grain purchases, and increased need for supplemental foods. Households in more distant areas showed no changes or had increases in livestock populations because their herds had fewer animals with which to compete for forage. When households neighboring the grazing reserve were compensated for the lease of the lands they had used, they prospered. We describe some benefits and limitations of the agent-based approach.
Key words: Agent-based modeling; DECUMA household model; ecosystem services; livestock; Maasai; mobility; pastoral decision making; SAVANNA ecosystem model
For millennia semi-arid and arid lands have supported wildlife, livestock, and the people who rely on those animals. Decisions by pastoral people have been influenced by factors such as forage quality and quantity, water sources, and fuel wood availability—which are more broadly labeled as ecosystem services (Daily 2000, Daily and Matson 2008)—while contending with stressors such as droughts, livestock raids, and changing markets. During the twentieth century, the ability of pastoral people to respond to stressors began to erode. Fragmentation and changes in land use and land tenure limited movements by wild and domestic ungulates (Behnke et al. 1993, FAO 2001) and reduced animal forage (Boone et al. 2005), and human and wildlife conflicts have increased (Browne-Nuñez and Jonker 2008). In Africa, human population growth severely limited the capacity of pastoralists to respond to stressors (Fratkin and Smith 2005). The twenty-first century will bring to semi-arid and arid systems unprecedented climate change, at least within human history (IPCC 2007), thus further stressing these systems.
What are the best ways for pastoral people to respond to new stressors? A main means of addressing the question quantitatively is through scenario analyses using computer simulations (Galvin et al. 2004, 2006, Peck 2004). Baseline simulation results representing a stylized version of current conditions are compared to simulation results where ecosystem services have been altered, or where the adaptive capacity of households has changed.
In past research we simulated how changes in ecosystem services would affect pastoral household well-being (Thornton et al. 2003, 2006; Galvin et al. 2004, 2006; Bulte et al. 2008). However, the household model used, called PHEWS, was population based, meaning that households were placed into a few groups using a classification we assigned (i.e., wealthy livestock owner; poor livestock and business owner). Three main restrictions arose from this approach. First, the top-down assignment of households yielded results at a given scale, and responses at finer scales could not be explored. Second, households could not shift between categories as their conditions changed, such as changing wealth levels. Third, and perhaps most importantly, because the finest representation of households were as members of these classification groups, households had no spatial location, and so the decision making of households could not be linked to conditions in their environments.
In this work, we sought a bottom-up organization of households. We chose a spatially explicit agent-based approach, where autonomous interacting agents make decisions based on their environmental and socioeconomic conditions and on a set of rules or processes (Epstein 1999, Bonabeau 2002, Evans and Manson 2007). Specifically, we chose an empirical agent-based approach (Janssen and Ostrom 2006), where agents were households simulated to represent stylized real responses. Existing models of households in rangelands were simpler than we required, dealt primarily with land cover change, were not spatially explicit, were not appropriate to link to our ecosystem model, or did not track food energy and monetary flows (e.g., Parker et al. 2003, Evans and Kelley 2004, Castella et al. 2005, Kuznar and Sedlmeyer 2005, Gross et al. 2006, Milner-Gulland et al. 2006, Cioffi-Revilla et al. 2008). We therefore constructed a suitable model.
In our study, we used two linked models to represent coupled natural and human systems, a long-established ecosystem model, and the new agent-based model. Our primary purpose is to introduce our approach to simulation of coupled natural and human systems, our agent-based model, and the methods we used to link that model to an ecosystem model. The agent-based model and the linkage of the two models are novel contributions, and are described in more detail than the ecosystem model. The household model is described using the ODD protocol (Grimm et al. 2006; Polhill et al. 2008). We reviewed a baseline simulation and assessed results by comparing them to observed patterns. We contrasted results from a baseline simulation with those from a simulation that included a new drought. Lastly, we used the coupled agent-based household and ecosystem models to simulate a scenario (Thornton et al. 2006) that we explored using the population-based household model. In that scenario, a high-elevation key resource area now used as a grazing reserve was converted to rain-fed agriculture. We discuss some advantages and disadvantages of each modeling approach, and interpret our results in light of Maasai household well-being.
Kajiado District is a semi-arid region in southwestern Kenya (Fig. 1; 36° 0’ E to 37° 55’ E, 1° 1’ S to 3° 3’ S) which is inhabited by Maasai pastoralists and others. Our study area is the southeastern half of the district (Fig. 1), an area of 10 746 km2. Amboseli National Park is near the center of the study area, and on the eastern border is the West Chyulu Game Conservation Area (Ole Katampoi et al. 1990). Precipitation is variable over space and time, but sums to between 400 and 800 mm annually, with the higher amounts occurring on slopes. The landscape supports diverse grasslands, extensive bushlands, five large swamps, and scattered Acacia woodlands, with some forests. Diverse wildlife populations inhabit the park during the dry season, and move into neighboring grazing lands in the wet season.
We estimated there were 52 000 people in the study area in 2002 (Thornton et al. 2006). Livestock raising remains the primary contributor to livelihoods, but Maasai have diversified (BurnSilver 2007). In essence most residents are agro-pastoralists, doing high-risk rain-fed cultivation of maize and beans on small plots. Others do more intensive irrigated cultivation in or adjacent to the swamps. Household members own small-scale businesses, and wages comprise a major income source for some Maasai households. The district remains monetarily poor (Government of Kenya 2003), with rates of poverty varying from 11 to 68% of the population (Thornton et al. 2006); people were classified as in poverty if a person earned <1239 Kenyan shillings (KSh) per adult per month, which is roughly equivalent to US$16.
BurnSilver (2007) and Worden (2007) surveyed six Maasai communities that differed in their history of land subdivision, and those surveys inform our household simulation model. We sought to model all households within the study area, but household survey data (Fig. 1) were for: (1) Osilalei Group Ranch; (2) Eselenkei study area, in the northern portion of the group ranch with that name; (3) Linkisim, in the southern part of that group ranch; (4) Emeshenani, which is in Olgulului/Lolorashi Group Ranch and abuts Amboseli National Park; and (5) northern and (6) southern portions of Imbirikani Group Ranch.
To link the mutual influences between households and ecosystem services through space and time required a spatially explicit ecosystem model, a household model appropriate for linkage to the ecosystem model, and information sufficient to allow the models to be parameterized and assessed. As an ecosystem model, we used SAVANNA, which has been useful in past work in the study area (Boone et al. 2005, Thornton et al. 2006, Boone 2007, Boone and Wang 2007). We constructed DECUMA (DEcisions under Conditions of Uncertainty by Modeled Agents) as a spatially explicit household model.
SAVANNA ecosystem model
Development of the SAVANNA model began more than 20 years ago, while its author M. Coughenour worked in the Turkana region of Kenya (Coughenour 1985). Since that time, the model has been updated and applied throughout the world (e.g., Coughenour 1992, Eastman et al. 2001, Christensen et al. 2004, Boone et al. 2006, Thornton et al. 2006, Boone and Wang 2007). SAVANNA is a series of connected FORTRAN modules that simulate ecosystem processes through time in a spatially explicit way. Landscapes are divided into cells, and digitized maps inform SAVANNA of the attributes of cells. Weather data for stations are used to interpolate monthly temperature and precipitation surfaces for the study area. During a simulation, plant functional groups within each landscape cell compete for light, nutrients, water, and space. During any time-step, plants in functional groups may grow, may reproduce, and may die, either through baseline death rates or stresses such as drought and extreme temperatures. The death of plants in one functional group may allow another group to expand its proportion of cover on a cell. Wild herbivores are represented in SAVANNA as populations. Herbivores feed on plants according to diets that are specified. Wild herbivores gain energy from the food they consume, and expend energy through basal metabolism, travel, gestation, and lactation. Excess energy is put to weight gain, and an energy deficit leads to weight loss. Wild animal population dynamics are simulated.
SAVANNA simulates ecosystems using a weekly time-step, with spatial and temporal summations produced each month. Simulations span from 10 to 50 years or more, and can simulate small areas or areas up to many thousands of square kilometers. More detail is available in Ellis and Coughenour (1998) and Boone et al. (2005).
DECUMA household model
The household model description follows the ODD protocol for describing agent-based models (Grimm et al. 2006, Polhill et al. 2008), which includes seven regular elements that provide an overview, design concepts, and details of the model.
DECUMA simulates decision making and behaviors by pastoral household heads as they relate to ecosystem services. Measures reflecting the well-being of household members, such as livestock dynamics and holdings, energy flows, and cash flows, are tracked. DECUMA links to an ecosystem model that quantifies ecosystem services (e.g., forage availability) and can simulate effects of grazing by livestock on services.
State variables and scales
The attributes of individual households are defined by state variables such as number of members, livestock holdings, incomes and expenses, and geographic location (Table 1, Appendix 1 provides an example based on a household interviewed). Decision making is influenced by a series of parameters common across households that capture attributes such as energy requirements, and adult-equivalency values, prices of animals bought and sold, and parameters reflecting the likelihood of seasonal movement by households (Appendix 2, with entries based on the literature or averaged responses from households surveyed). Households interact with other households through competition for grazing resources and by gifting of livestock. A weekly time-step is used to simulate livestock dynamics and energy acquisition, and a monthly time-step is used for birthing and aging of livestock herds, and for household decision making. There is no intrinsic spatial scale associated with DECUMA, as that is provided by the ecosystem model to which the agent-based model is joined. In the application to southeastern Kajiado District, Kenya, the gridded landscape cells represented areas 2.5 x 2.5 km, with 10 746 km2 simulated. Results may be summarized for all households within arbitrarily defined subareas, or for the entire area simulated. Spatial and temporal results are produced monthly, and simulations typically represent from about 10 to 50 years. Here, simulations spanned 24 years.
Process overview and scheduling
Process may be grouped into four broad categories, those that simulate (1) livestock distribution and dynamics, (2) household decision making and flows, (3) initialization, and (4) input/output. Here we focus on the first two. Livestock processes include: distribution of animals of simulated species, based on forage availability provided by the ecosystem model and on the locations of households and rules of access; energy acquisition, based on the amount of forage acquired given the distribution of the livestock; energy use; weight change, which is based on the difference between energy acquired and energy used; and population dynamics, with birth rates and death rates related to a ratio of current and expected body mass.
Household modeling includes the following processes: energy flows, where caloric gains from foods eaten are tallied and compared to energy needs; cash flows, which includes regularly scheduled income and expenses, as well as short-term sales or purchases; crop harvest; livestock trading; a calculation of cash needs three months into the future, which is used in decisions about livestock trading; gifting; and mobility of camps, where herders decide whether to move their temporary camps.
Regarding scheduling, after the model initializes, a monthly cycle begins. Herders weigh the ecosystem services, especially forage quality and quantity, at their current location and at a set of randomly selected alternate locations, and may decide to move if the anticipated benefits (energy acquisition) outweigh the costs (e.g., travel costs, distance to water, being away from the home group ranch). A weekly cycle of livestock grazing is then modeled. Livestock are distributed on the landscape based on habitat suitability, then the energy the animals acquire from grazing in that distribution is summarized, and livestock status is updated. The monthly cycle then resumes. The energy acquired by livestock and energy costs are used to model changes in body mass. Condition indices are updated, which are scores from 0 to 1 that compare simulated to expected body masses. Mortality is then simulated, with condition indices of animals having an effect on their mortality rates. At the appropriate month females give birth, again with the condition of animals influencing rates, and at the end of the year livestock are aged.
Crops are harvested if the month simulated matches the month assigned for the crop. The primary sources of household income (e.g., crop sales, milk sales, wages) and expenses (e.g., food, schooling, veterinary care) for each household are reckoned to yield monetary flows and holdings. The model calculates the money each family may need in the following three months, based on predictable expenses. That information is used by the household to decide if a cow, goat, or sheep should be sold (Thornton et al. 2003). The model then tracks food energy acquired by each household, including from milk, home-grown maize, edible dead animals, meat eaten during ceremonies, and sugar in tea. If a deficit in energy still exists and the household members can afford to, they buy maize. If a shortfall still remains and more livestock cannot be sold, it is filled through supplemental relief from neighbors or international aid agencies (Thornton et al. 2003). If household members have ample money in reserve, they may buy an animal. Animals are bought from, or sold into, an unlimited pool of animals outside the simulated area. Finally, a few animals of a given species may be given by wealthy families to families who have lost their herd of that species (Huysentruyt et al. 2009). The model then continues simulating the next month.
The main processes and connections in the model are shown in Fig. 2, including processes simulated in the ecosystem model. Appendix 3 provides the main controlling program of DECUMA, showing the explicit scheduling in the model, with each line annotated.
Simulation flow. A major design concept was the need to link DECUMA with ecosystem models. In linking with SAVANNA, soils, vegetation, and wildlife are simulated as in past applications (Coughenour and Singer 1996). For livestock, SAVANNA and DECUMA share information weekly in the following process (Fig. 3):
Prediction. Agents make a form of prediction in two ways in DECUMA. First, households know their scheduled incomes and expenses for the coming months (e.g., Appendix 1). Households anticipate known cash needs three months into the future, and use that information in livestock trading; if more money is needed over the next three months than is held, they are more likely to sell an animal. The second form of prediction of future conditions is through use of long-term habitat suitability surfaces. These surfaces are the average habitat suitability of areas for each species throughout the simulation. They capture long-term forage availability expectations and are used by agents when deciding whether or not to move.
Sensing. Household herders sense the habitat suitability within a distance defined as the grazing orbit around their current location (e.g., 10 km). Sensing also occurs over the entire area when herders consider moving their temporary camps, i.e., when the benefits from moving to 10 sites selected randomly from the area are weighed. This long-range sensing reflects the sharing of news about grazing conditions, which is a common pastime in the community.
Stochasticity. The order in which herders decide where to graze their animals is relevant, because those who decide first have the best grazing areas to select from. Detailed political and power relationships are not available to us, and so we used a newly randomized order of selection each week (as in Milner-Gulland et al. 2006).
Observation. For each simulation, DECUMA produces tabular files that store attributes averaged across households. Another tabular file saves monthly data for individual households, i.e., for a subset of households a user selects. The model also produces spatial data, which includes permanent and temporary camp locations each month, habitat quality for each livestock species, and livestock distributions. Some output uses standardized measures for livestock and humans; tropical livestock units (TLU), with one unit equal to 250 kg body mass of livestock; and adult equivalents (AE), where adult males were assigned an AE of 1, and adult females and younger people were assigned smaller values (see Appendix 2). A graphical user interface created using Visual Basic 6 (Microsoft Corp., Redmond, Washington, USA) runs the linked DECUMA and SAVANNA models, and creates charts and maps of user-selected output as a simulation progresses.
Initial attributes for households were set using the conditions of the surveyed household that was geographically nearest to the location selected for the new house and in the same region. Attributes initialized for each household are summarized in Table 1 and an example is provided in Appendix 1. To avoid a long period of unstable responses at the start of each simulation, a “spin-up” simulation of 60 years was made for all the households, and then their conditions were stored to a computer file. This spin-up used randomized years of weather data from 1973 to 2002. During subsequent simulations, this file was read, and conditions for households were set to those in the file. The initial conditions of the surveyed households were then reset to the observed values, and simulations commenced.
DECUMA reads a series of maps and parameter files that describe the study area and household attributes, including maps delineating the study area, household densities, slope, distances to water sources, and the subareas of interest. Appendices 1 and 2 provide example parameter files. Three files provide age distribution, energy, and population parameters for each livestock species simulated. DECUMA does not use other dynamic input directly, but the ecosystem model to which DECUMA is linked uses dynamic precipitation and temperature data.
DECUMA is composed of a series of submodels programmed in FORTRAN 95. The primary submodels are described more fully in Appendix 4.
SAVANNA–DECUMA modeling in southern Kajiado
SAVANNA and DECUMA were parameterized to emulate conditions during the period when household interviews were gathered (1999 to 2000), to the degree possible. Seven plant functional groups are represented (i.e., palatable grass, palatable forbs, unpalatable herbs, swamps, palatable shrubs, unpalatable shrubs, and woodlands, as in Boone et al. (2005)). Six wild herbivore functional groups are included (wildebeest (Connochaetes taurinus), zebra (Equus quagga), African buffalo (Syncerus caffer), grazing antelope, browsing antelope, and elephant (Loxodonta africana)). Species included in the antelope groups and example citations used in parameterizing the model are in Boone (2005). In general, SAVANNA–DECUMA represents land cover types spatially, but cultivation is non-spatial (i.e., area cultivated is an attribute of households; Table 1, Appendix 1). Cultivation by Maasai households is at relatively small scales (i.e., <1 ha on average), but landscapes were represented by 2.5 x 2.5-km cells (i.e., 6.25 km2 or 625 ha), making a spatial representation of cultivation impractical. Simulations reported here use precipitation and temperature from 1980 to 2003, with years labeled 1 to 24 in the figures.
We simulated 3820 households (Thornton et al. 2006). Detailed survey data for 184 of those households were available (BurnSilver 2007, Worden 2007). Household densities were mapped using census and ancillary data (Thornton et al. 2006). We distributed the 3820 households by pseudo-randomly selecting locations based on the densities in that map. The households were then initialized as described above.
Assessing results from integrative simulations such as this is particularly difficult (Thornton et al. 2003); the utility of the results to researchers and stakeholders becomes paramount (Rykiel 1996). We would prefer to have an extensive, unique set of household survey data to compare with, but those data are not available. Our best observed data are the household surveys used to initialize our model. With six subareas in the simulation, and given the variability between households, we chose not to reserve some of the data for assessment. Pattern-oriented assessment (Grimm et al. 2005) suggests that the agreement of results to multiple patterns at different scales can help in assessment. We therefore compare our simulated results over time to patterns in the household survey data and to community-level patterns (e.g., poverty rates).
In the first scenario, we looked at the effects of a 2-yr drought on households and the ecosystem. Drought was used because its primary effects on semi-arid and arid landscapes are widely known. In the weather data, we selected a period of typical rainfall (1985 to 1986) and decreased precipitation to equal the mean of precipitation (550 mm) minus twice its standard deviation (150 mm; following Galvin et al. 2004). In our second scenario, we emulated a setting where access by the Maasai to a key resource, the Chyulu Hills dry season grazing reserve, was lost and households were not compensated, such as may occur if the Kenyan government chooses to use those lands for wheat production. Our third scenario is related, but with the Maasai leasing their lands for wheat production, and receiving compensation (as in Thompson and Homewood 2002).
As in reality, changes in livestock populations and household responses are sensitive to climatic patterns. Our drought scenario focuses on the observed and modified weather pattern, and so we used those data in two simulations. For our scenarios regarding access in the Chyulu Hills, climatic variability was not our focus, so we conducted 50 simulations for each scenario, using a unique random ordering to annual weather data in each simulation, 25 for the baseline model, and 25 with access denied.
In our third scenario, the Chyulu Hills, were leased from members of the neighboring group ranches. We appreciate Thompson and Homewood’s (2002) message that benefits to group ranch members from external sources are not evenly distributed because of power and access imbalances and graft. However, those relationships are notoriously opaque, and were not available to us. Here, each member received the same benefit. In ranches in Narok District, Kenya where lands were leased to wheat cultivators, ranch members made on average US$25 ac-1 yr-1 (Thompson and Homewood 2002:129). The Chyulu Hills have a similar agro-climatic potential as the lands in Narok (Ole Katampoi et al. 1990, drawing on Braun 1980, Thompson and Homewood 2002), and so we adopted this value. Based on the area of the Chyulu grazing reserve (81 250 ha or 200 773 ac) and numbers of households who were members in the two group ranches that abutted Chyulu, we calculated monthly income to each Imbirikani and Kuku household would be 2490 Ksh, or US$34.58/month, using the exchange rate of 72 Ksh/US$.
Average responses for the 184 focal households from DECUMA were calculated. For the scenarios regarding changing access to the Chyulu Hills, responses from the baseline model using one of the randomized weather files were subtracted from the matching simulation where access was altered. This yielded differences in responses with the expected sign from the paired simulations. We then calculated averages and standard errors using SYSTAT Ver. 11 (2004; Chicago, Illinois, USA) and created figures. One-sample t-tests were used to compare mean differences in responses across our 184 households to no change in responses (i.e., mean = 0). Bonferroni adjustments were made for multiple comparisons. Our results emphasize DECUMA output, in order to demonstrate the new model.
Bounded baseline responses are a type of assessment, in so far as keeping the numerous responses in DECUMA concurrently reasonably bounded while maintaining responsiveness to stressors such as drought is non-trivial. Comparing our baseline results (Fig. 4) to observed patterns at the scale of households, we simulated slightly fewer livestock per person than were observed among the 184 households surveyed (BurnSilver 2007:48), and with greater variability across households (5.6 TLUs/AE, SD 11.4 versus 6.3 TLUs/AE, SD 6.7). Total livestock populations were stable, but there was a gradual increase in the amount of supplemental relief required by households, a decline in milk and meat energy acquired, a slight increase in the number of animals sold over time, and an increase in energy that was purchased (Fig. 4). The number of movements to temporary camps was in line with observed rates, and greater in 2000, a year of drought, than in 1999 (observed, 2.2 and 2.7 in 1999 and 2000; simulated, 2.6 and 2.7). Income across the households was in close agreement, with US$1583 earned on average in surveys (BurnSilver 2007:53) and US$1572 in simulation. At a broader scale, poverty rates in Kajiado are high (i.e., incomes below US$16 month-1 AE-1) (reviewed in Thornton et al. 2006). In the base simulation, average monthly income per AE is US$9.70, with high variation (SD US$5.32).
Effects of drought
In the scenario results, effects of drought on Maasai livestock and well-being are longer-lived than may be anticipated (Fig. 4). This does not reflect rangeland degradation, although that occurs in the short term, but rather the steps non-wealthy Maasai must take to meet their caloric needs. The severe drought decreased livestock numbers (Fig. 4a), which reduced animal-source foods for household members (Fig. 4c and d). Drought also eliminated rain-fed maize production for the households, thus further reducing food security. Households without monetary stores had to then sell livestock (Fig. 4b) to purchase grain (Fig. 4e), and the shortfall was made up with supplemental gifted relief (Fig. 4f). The sale of livestock in turn led to less food available in the next month, and the process continued as a positive feedback loop, yielding a downward spiral sometimes seen in reality (Rutten 1992, Boone et al. 2005). These linked responses highlight the interconnectedness of the DECUMA model. Individual household responses may be compared given the agent-based focus. For example, shifts in cattle holdings (Fig. 5) and investigation of individual household responses confirm that although most households lost animals in drought some increased their herds, due to reduced competition for forage.
Changes in access and compensation
Loss of access to the Chyulu Hills grazing reserve by Maasai herders caused a decline in the number of livestock per person in the Imbirikani study areas, Emeshenani, and Linkisim (Fig. 6a and b, left column). These declines may appear modest, but represent about a 25% decline in numbers of animals owned by households in Northern Imbirikani, which would be severe for households already experiencing food insecurity. In Eselenkei, livestock populations declined initially, but later increased when competition for forage lessened as herds in the areas closer to the Chyulus declined. Herders from near the Chyulu Hills had fewer animals they could bring to Eselenkei during the dry season, thus reducing competition and benefiting the local livestock. These responses are tempered because herders in Eselenkei and Osilalei lost access to the Chyulus during severe drought, as did the other households in the study area. When households in Imbirikani Group Ranch were compensated for their loss of access to the Chyulu Hills, they prospered. Households in Imbirikani purchased additional animals, but also avoided having to sell animals to purchase grain (Fig. 6a, right). Livestock populations in the remaining areas did not change markedly (Fig. 6a and b). This further demonstrates the value of place-based simulation of agents, where those living near the Chyulus could increase their herds using compensation, but they could not support an unlimited number of animals on the wet season forage available around their permanent households. Too few animals were moved to Eselenkei and the other more distant areas in the dry season to cause population declines in the herds of the households that lived there. Changes in other measures of household well-being as access to the Chyulu Hills was lost are summarized in Table 2, with significant differences noted.
Our results demonstrate that the linked DECUMA and SAVANNA models simulate the coupled human and natural Kajiado ecosystem reasonably, with baseline responses for the households for which we have survey data remaining reasonable throughout simulations. Changes in livestock were less dramatic and more locally variable in these analyses than in the parallel analyses of Thornton et al. (2006). In that work, livestock were modeled as populations, and each month livestock were redistributed on the landscape. In resource poor months, many thousands of animals would be placed in the Chyulu Hills, and then placed a hundred kilometers away the following month. Livestock thereby made ready use of the grazing reserve without travel costs. In the agent-based approach, livestock were associated with specific places on the landscape. During typical dry seasons, households far from the Chyulu Hills did not travel to them, as in reality (BurnSilver 2007), and so those livestock were only affected by the change in access through interactions with animals from areas closer to the Chyulu Hills.
Subdividing a portion of the Chyulus for use by sedentarized Maasai for rain-fed agriculture has been discussed by community members, and simulated (Boone et al. 2006). A borehole pipeline has been constructed into the core of the grazing area, which if managed poorly would allow overgrazing in the reserve, as simulated in Galvin et al. (2008). Collectively, our results emphasize the importance of maintaining access to the Chyulu Hills for Maasai pastoralists. Novel results emerge here, given the place-based nature of the agent-based approach. The loss of access to the grazing reserve caused households closest to the Chyulus to lose livestock, but households more distant from the Chyulus gained livestock, because there are fewer livestock immigrating during the dry season and their herds had more forage per animal. The opposite response did not occur when people near the Chyulus were compensated for their loss of access. Household heads purchased more livestock, but they still had to support those animals near their permanent households in the wet season, and the number that could survive was limited. That number was insufficient to cause households more distant from the Chyulus to lose animals due to the increased competition for forage.
Two effects contribute to trends in responses in our baseline model (Fig. 4). First, Kajiado pastoralists return their herds to their permanent homes during the wet season (BurnSilver 2007), but in the simulation some permanent households occupy areas that cannot provide sufficient forage. Households in these locations may lose animals in most years, and be restocked through gifts. Second, families with members earning salaries may purchase many animals. Some families lose animals and other families purchase animals, but overall the total livestock numbers remain stable over the long term. However, poor households far outnumber wealthy households and so the median household response was a gradual decline in livestock numbers. Indeed, the portion of households owning half the livestock was 12.5% in surveys and 8% in the base simulation.
The ability to summarize responses at individual-to-population scales leads to a main challenge of our approach. In the PHEWS model, the numbers of control parameters and output categories were similar. In general, controls could be adjusted to alter responses more-or-less directly. In agent-based DECUMA, the number of control parameters is far less than the responses. For example, here about 11 460 herds were being simulated. The same types of adjustments to parameters are made in DECUMA as in PHEWS. But upon summation of the results, if some households are selling far more than they should and some selling far less, that cannot be adjusted directly. Instead, one must consider why the differences may be occurring and make adjustments through a synthetic systems approach. This way of modeling is appropriate, but the process of parameterizing the coupled simulations can be iterative and complex.
Our simulations are not intended to represent conditions 24 years into the future. Too many changes to the linked systems will occur in that time to make such predictions possible. However, our results do quantify household well-being over time if a single change in access occurs, for example. Moreover, the caveats associated with SAVANNA modeling apply here as well (see Boone et al. 2005, Boone 2007).
With DECUMA linked with the SAVANNA model, changes in ecosystem services influenced the decisions made by household members, which in turn influenced ecosystem services. Linking complex models brings challenges, but this approach increases the potential for emergent responses and secondary interactions to be considered. Using DECUMA, responses may be summarized at different scales, attributes of individual households are known and can change over a simulation, and pastoralists have permanent household locations and temporary camps with environments that influence their decision making.
Responses to this article are invited. If accepted for publication, your response will be hyperlinked to the article. To submit a response, follow this link. To read responses already accepted, follow this link.
ACKNOWLEDGMENTSWe thank three anonymous reviewers and the Subject Editor for providing comments that helped us to improve the manuscript. Our study of resiliency in livestock-owning households and the development of DECUMA is supported by the National Science Foundation, grant SES-0527481 to Galvin et al., with additional support from grants BCS-0822752 and BCS-0624315.
|Home | Archives | About | Login | Submissions | Subscribe | Contact | Search|