Kennedy, W.G., Cotla, C.R., Gulden, T., Coletti, M, and Cioffi-Revilla, C. (2014) Towards
Validating a Model of Households and Societies of East Africa. Advances in Computational
Social Science: The Fourth World Congress, Chapter 20, pp 315-328, S.H. Chen, I. Terano,
H. Yamamoto, C.C. Tai (Eds.) Springer.
Towards Validating a Model of Households and
Societies in East Africa
William G. Kennedy, Chenna Reddy Cotla, Tim Gulden, Mark Coletti, Claudio
Cioffi-Revilla
Abstract One of the major challenges of social simulations is the validation of the
models. When modeling societies, where experimentation is not practical or ethical,
validation of models is inherently difficult. However, one of the significant strengths
of the Agent-Based Modeling (ABM) approach is that the approach begins with
the implementation of a theory of behavior for relatively low-level agents and then
produces high-level behaviors emerging from the low-level theorys implementation.
Our ABM model of societies is based on modeling the decision making of rural
households in a 1,600 km (1,000 mile) square around Lake Victoria in East Africa.
We report on the first validation our model of households making their living on
a daily basis by comparing resulting activities against societal data collected by
anthropologists.
Key words: Agent-Based Modeling, household decision-making, East Africa
1 Introduction
One of the major challenges of Agent-Based Modeling (ABM) is the validation of a
model. When modeling societies where experimentation is not practical or ethical,
validation of models is inherently difficult. However, one of the significant strengths
of the ABM approach is the faithful implementation of a theory of behavior for relatively low-level agents and their associated environmental dynamics and then observing high-level behaviors emerging from the low-level theorys implementation.
That is the approach we have taken toward validating our model.
A team of scientists at George Mason University and at Human Relations Area
Files (HRAF) at Yale University has been working a few years on an agent-based
Center for Social Complexity, Krasnow Institute for Advanced Study
George Mason University, Fairfax, Virginia 22030, USA e-mail: wkennedy, ccotla, tgulden, mcoletti, ccioffi@gmu.edu
1
2
Kennedy et al.
model of a large area of East Africa, including validation-related fieldwork in 2010.
The purpose of this project is to answer research questions on social dynamics, such
as internal conflict and responses to natural disasters and humanitarian relief. This
follows earlier validation efforts in the same project on other components [1–4].
This work reported here is the progress of our efforts to validate our model of
household subsistence decision-making. We will discuss our validation efforts organized by the layers of our model up to the model of household decision-making.
Later papers will describe further work with this model and address the reaction
of people suffering from disasters that affect their subsistence activities, such as a
prolonged drought.
Our model includes detailed representations of the environment, specifically land
types, water supplies, and weather. From these we have modeled vegetation growth
for grazing of domestic herds and farming activities. We have modeled the people in
the region at the household level, and households manage herds, farming, and labor
activities. These subsistence activities are modeled down to actions taken on a daily
basis, such as deciding when to plant, when to harvest, and where to move the herd
each day.
It is relatively easy to compare modeled vegetation to actual data on vegetation.
To test the macro-level performance of our model of household behavior, we use anthropological data from local ethnographies, such as primary sources from HRAF
and secondary sources from extant literature (e.g., [5–7]). Anthropologists have catalogued approximately 135 different ethnic groups in the modeled area, including
data on how each the people of each culture makes their living. Therefore, by implementing theoretical decisions at the household level driven by the environmental
conditions, we can see if the simulated results match anthropological data. We report on the validation of our household model against the anthropological data for
the region, which will provide insights into model validation for models of societies.
2 RiftLand model overview
Our model of the Rift Valley of East Africa surrounds Lake Victoria. Our model,
called RiftLand, was developed using the MASON system [8, 9] and represents
the area and the modeled actors at a scale appropriate for our research questions.
The subject area is show in Fig. 1. As an agent-based model, we attempt to model
the environment and agents at a relatively low level and allow their interactions
to produce macro-level results. The environment is represented at the one square
kilometer level throughout the 1,600 by 1,600 km area. The time scale has one
step of the model representing one day. People are modeled at the household level,
keeping track of the number of individuals in each household. Our model runs are
typically over several years of simulated time, or thousands of days. Major model
components and behaviors are detailed below to a level necessary to discuss our
validation efforts. A more detailed description of the model to support replication
of our work will be published later [10].
Towards Validating Households
3
2.1 Reference Data for Validation
Many anthropologists have lived and studied the people of the Rift Valley and their
observations have been collected into an atlas of the people of this region and how
they make their living. The atlas was published by G.P. Murdock as a series of
journal articles [11, 12] in the journal Ethnology in the late 1960s. This information
has been updated by HRAF. The 135 HRAF-coded ethnographic cultures of the Rift
Valley region are shown in Fig. 2.
2.2 Modeling the Environment
The subject area includes several biomes, cultures, and polities. We represented the
1,600 by 1,600 km area as date representing each 1 km2 parcel of land or water. This
level of resolution supports reasonably accurate representation of the landscape, political borders, and locations of the people. We, of course, did not model every rock
and blade of grass, but did model major factors, actors, and interactions affecting
the macro-level behavior of interest. We represent differences in land use by different types of parcels. Parcels can be all water, either saltwater (in the Indian Ocean)
or freshwater (1 km2 parcels of Lake Victoria, other named lakes, and some major
rivers). On land, we differentiate urban areas, forests, parks, and open parcels available for grazing or farming. The major difference is the impact on vegetation and
people.
Given the importance of water in this region of the world, we carefully modeled
water sources in the area. The easy part is placing water sources along large rivers
and around freshwater lakes. Using the elevation information for the region, we also
!
Fig. 1 The RiftLand model area and three test regions identified. Sources: NASA (left) and RiftLand model output with different land/water types (right)
4
Kennedy et al.
placed water sources in the lowest areas to represent seasonal water sources and
man-made watering holes. Using available data on the number of water sources in
some areas, we are able to appropriately populate a given region with water sources.
2.3 Modeling Vegetation Growth
Vegetation growth is very important to local populations through its impact on both
agricultural and pastoral activities. We modeled vegetation based on three factors:
local rainfall, land fertility, and remotely sensed data on the resulting vegetation
over time. Using rainfall data collected frequently over large parts of this area, we
developed approximate 50 weather cells covering approximately 30 km2 regions.
For each region and covered parcels, we generate a rainfall amount each day, which
is used to determine the vegetation produced on each parcel.
The amount of vegetation produced on a parcel is computed using a logistic
equation with rainfall as input. We use Normalized Difference Vegetation Index
(NDVI) [13], which provides daily vegetation data based on remote sensing, to val-
! 2 Cultural Map of Rift Valley Region, based on HRAF-coded ethnographic cultures. Source:
Fig.
HRAF.
Towards Validating Households
5
idate results. We then determine appropriate land fertility values for each parcel as
the residue of a linear regression model of rainfall and NDVI data, based on NDVI
as reference vegetation, our vegetation growth model, and rainfall data. These factors together allow the simulation to grow vegetation based on the lands fertility and
rainfall for both agriculture and pastoral modeling purposes.
2.4 Modeling Herds
Domesticated animals are modeled as generic tropical livestock units (TLUs), which
represents 1.0 head of cattle, 0.7 camels, 11 sheep, 10 goats, and other numbers of
other animals [6]. Using this abstract unit, we did not need to differentiate the types
of animals in a herd. As a herd of TLUs, we modeled the herds need for water and
vegetation. These physiological parameters allowed us to model the daily intake
needs, current levels, and general health of the herd. We also assumed a birth rate
and death rate when these physiological needs were not met, resulting in each herd
being modeled as a number of TLUs with daily needs and health status as a function
of water and vegetation.
3 Household Modeling
The model keeps track of the number of people engaged in different daily subsistence activities, although people are modeled as households who make decisions.
The primary activities of households are farming and herding. Family members are
also modeled as engaged in wage labor, bringing cash to their household. Household activities are semi-independent subsistence activities. Ideally, each would be
self-sufficient, but the model recognizes the activities as part of one household and,
if necessary, the people of one activity can live on resources of another. After describing the activities that household members could engage in, we describe how
they decide.
3.1 Farming Activity
Only the key parts of farming were modeled. These included planting a specified
amount of land and harvesting its crop. Part of the households population was assigned to farming. This was envisioned to include small children and to occur on the
households farm, one of the parcels suitable for grazing or farming. Those engaged
in herding, which could be a relatively independent activity, were expected to be
at the households farm to provide labor from two weeks before planting until after
harvest.
6
Kennedy et al.
Farming involves deciding when to plant, how much land to plant, and how long
the growing season will be. Vegetation growth was then modeled daily over a growing season of approximately 90 days. Planting (i.e., starting from zero vegetation),
occurs on a specific day and, after the adjustable growing season, the crop is harvested. We included a farming productivity factor that increases the yield on tended
farmland over what would grow wild. However, we did not specify the crop type.
Vegetation was modeled as producing a yield in units of kilograms of dry matter,
for which we had production and consumption data.
The history of rainfall on the farm indicated the best month to plant. Based on the
regions weather pattern, there could be one or two best times to plant and the farm
was modeled to have one or two growing seasons. A week before the scheduled
harvest, the farmer makes an assessment of the crops potential yield. The farms
next growing season is then extended or shortened based on comparing the potential
harvest a week before the actual harvest.
At harvest, the crops yield is added to the farms grain store. The farms store is
the source of daily food for farmers and other household members, if needed. After
harvest, those assigned to the herding activity take their TLUs and go to open land
to graze the herd until next season.
3.2 Herding Activity
Herding is concerned with where to move a herd each day. To make that decision,
we have employed a fast and frugal decision tree that prioritizes concerns for a need
to change watering holes, avoid conflict, and move herd to graze or to water the
herd. The decision-making process was described previously [3].
Herders make their daily living directly from animals. A set number of TLUs
is necessary to support a herding household. If the number of TLUs in the herd
decreases below the number necessary to support the herding household, then its
subsistence depends on whether the herd is in the field or at the households farm.
If at the farm, the unsupported herders can subsist from farm store or cash reserves.
If in the field, we presume they can find a living from other sources. If the herded
TLUs increase beyond what the herding household can manage and the herd has
returned to the farm, the animals are sold to increase the households cash assets.
With the general practice of sending teenage boys with the herd when there is a
household base on a farm, the people engaged in herding do not reproduce. However, if the household does not have a farm and the household is therefore moving
with the herd, normal reproductive activity takes place.
Towards Validating Households
7
3.3 Labor Activity
All households in the RiftLand area need cash for some obligations, such as paying
taxes and paying for schooling. We address this need by including labor activities
when households involve more than ten members. Household members involved
in labor activity generate cash for the household and are presumed to not need to
subsist through other household activities. Human resources in labor, however, can
be re-assigned to support farming or herding activities, if needed. Also, if herding
or farming activities are not self-sufficient, the unsupportable individuals will be
transferred to labor activities. However, if the fraction of the household involved
with labor becomes too great, those persons become internally displaced (so-called
IDPs) and are no longer part of the household.
3.4 Displaced Persons
Displacement is generated as an emergent phenomenon in the model and displaced
persons are part of the models rural household decision-making. The intent, in a
later version of the model, is that displaced people move to urban centers for subsistence and if their numbers exceed the capacity of the local urban center, they will
move to larger urban centers, possibly also moving across national borders officially
becoming refugees.
3.5 Household Decision Making
Each household makes decisions concerning human resources to apply to its subsistence activities. If all goes well, a household would have an operating farm, herd,
and enough members to have some members engaged in labor. On a daily basis
when the herd is at the farm, all activities share the resources necessary to meet daily
food needs and excesses become accumulated wealth. If an activity does not do well
enough to feed its assigned people for some period of time, some of the people are
re-assigned to another activity, as shown in Fig. 3. The reason for all re-assignments
is based on the success of the activity in providing food for the household; changes
in effectiveness of the activity results in transfers of household human resources
among activities. The lines indicating re-assignment from the farming and herding
activities to labor are different because the transfer to labor only occurs if the other
major activity cannot feed their assigned people. Farming activities are evaluated at
each planting and harvest, while herding activities are evaluated every four months.
If farming or herding activities fail completely, the household will try the failed
activity again after some time by re-employing laborers. If the primary activities of
farming and herding fail and too many are simply laborers, the household becomes
displaced. On the positive side, successful farming or herding will result in starting
8
Kennedy et al.
a new household with approximately half the original households assets. Successful
households divide their resources creating a new farm and herd as appropriate and,
if large enough, some labor. Our 1 km2 parcels can support several co-located farms
and herds. However, each parcel has a maximum carrying capacity.
4 Modeling Household Behavior: Preliminary Results
The RiftLand model has been run under a variety of conditions to demonstrate how
households make a living throughout the region. The expectation is that environmental factors will cause different subsistence strategies in different regions. We
have focused on three different regions to validate our model of household decisionmaking in the RiftLand area, since overall variability is exponentially (not just linearly) proportional to the number of individually validated regions: one valid region
is good, two are better, three are far better. Three regions studied are shown in Fig. 1
as white squares. In each of the three study areas, our simulations involved placing 1,000 households in the region and simulating their behavior over 3,650 days
(i.e., 10 years). Households are initialized with a number of persons divided nearly
evenly between farming and herding activities. If the household has greater than 10
persons, then approximately 10 percent are initialized as being engaged in labor to
generate cash for paying taxes, educational, or other needs. Households start with
both farming and herding, approximately balanced, and subsequently adjust human
resources applied to each activity based on their success in feeding the household.
The model actively manages the number of people involved in those two activities
!
Fig. 3 Re-assignments of Household Human Resources among Activities
Towards Validating Households
9
and only if there are more than 10 people in the household are any assigned to labor. The labor activity is mostly used as the placeholder for herders or farmers that
cannot be fed by their activity. Preliminary results are presented in the next three
subsections.
4.1 Survival in Semi-Arid Regions of Northeast Kenya (Region 1)
The first area tested is in the relatively dry northeast region of the modeled area.
We focused on an area 150 km by 150 km in northeast Kenya indicated as a white
square numbered 1 in Fig. 1 (right). This area, near the Bokhol Plain (500–1000
m elevation) in Wajir District is close to the Mandera region studied in previous
work [3, 4]. The area lacks major population centers (closest is Wajir, ca. 30,000
pop., 100 km away), water supplies, forests, or parklands. This is an arid, open
area with low population density (less than 10 inhab./km2 ) inhabited primarily by
Somali people. Fig. 4 is a plot of the simulations results. This plot presents the mix
of subsistence activities as the average of 1,000 households and shows those results
for 30 runs.
!
Herding
Farming
Fig. 4 Results of Household Decision Making in Region South of Mandera
10
Kennedy et al.
Fig. 4 shows average results of 1,000 households deciding how to make a living
over 10 years. The dramatic change after two years is the result of many household
agents deciding to give up their farming because it is not effective in feeding the
household members. The evaluations occur at harvest times and they are relatively
in sync early in the simulation and so many make the switch at the same time. After
more time, household re-assignments are more diverse due to local differences in the
weather and land fertility such that re-assignments at harvest are no longer visible
for the whole society. After the initial nearly synchronized major change and several
more years, the agents in this region settle on about a 90/10 preference for herding
over farming, in small households with very little labor activity.
The reference anthropological data for this region is contained in the Ethnographic Atlas [11, 12] as updated by HRAF. With the modern-day updates, the people of this region, the Bararetta [5], are described as PA, meaning they are both
pastoralists (herders) and agriculturalists (farmers), generally consistent with our
simulation results.
4.2 Survival in a Relatively Wet Area, Southwest Uganda (Region
2)
The second area tested is in the relatively wet region west of Lake Victoria. We focused on an area 100 km by 100 km in southwestern Uganda as indicated in Fig. 1
(right). The area size was selected to be primarily rural and to avoid major population centers, forests, or parklands. This region consists of open areas, available for
farming and grazing family herds, with relatively high population density (100-300
inhab./km2 ). Fig. 5 is a plot of the simulations resulting population in this region for
30 runs.
Simulation results show that the households much more gradually settle on their
mix of subsistence activities than in the previously discussed region. During each
of the first several years harvests, there were large shifts in activities before a few
years of consistent trends settled on approximately a 90/5/5 percent division between farming, herding, and labor.
Murdock’s 1967 atlas reported the Nyoro people [14] in this region are coded as
Ap, primarily agriculturalists with some pastoralist activity, specifically in the band
of 56–65 percent subsisting on agriculture and 16–25 percent subsistence on animal
husbandry. In our model, our agents are more heavily focused on agriculture.
4.3 Survival in Burundi (Region 3)
The third area tested was less arid (500-1000 mm annual precipitation), in Burundi,
southwest of Lake Victoria. We again focused on an area 100 km by 100 km to be
primarily rural and to avoid major population centers, forests, and parklands. This
Towards Validating Households
11
region of our model consists of some open areas available for farming and grazing
family herds. Fig. 6 is a plot of the simulations resulting survival strategies in this
region for 30 runs.
In this region, we see different behavior. Unlike the first region where most
households gave up farming after two years, here may do, but not as many as in the
first region. There is another shift, but much less dramatic after year eight. These are
likely caused by poor rainfall leading to farming being less productive for a short
period of time but better than herding over the long term. There is also an interaction between the amount of land involved in farming making herding less viable
in this region. However, overall farming was very good and produced households
large enough to support more labor activities than the other regions. The resulting
distribution was approximately 80/10/10 for farming, herding, and labor. The anthropological data for the Rundi people of this region are coded as Ap, but with
largest subsistence fraction being in the band of 46–55 percent in agriculture and
the second largest fraction being 26–35 percent in pastoral/herding.
!
Farming
Herding
Fig. 5 Results of Household Decision-Making in a Southwest (Ganda) Region of Uganda
12
Kennedy et al.
5 Progress in Riftland Model Validation
Our approach to validating this agent-based model is to start with micro-level models from which emerge societal-level behaviors that match available empirical (in
this case ethnographic) data. Our models of subsistence activities employed in East
Africa are based on available data, weather, TLUs, vegetation, and other socionatural features. Household agents adjust their assignment of human resources
based on the performance of their subsistence activities. Based on these inputs and
modeling dynamics, the aggregation of household decisions results in distributions
of farming, herding, and labor activities available for comparison with anthropological data. The available anthropological data coded subsistence activities in five
categories and we are attempting to have our model match the top two activities
(farming and herding). So far, our model has matched the primary subsistence activity in each regioneither as farming or herdingbut it does not yet match the approximate percentages reported for the three regions. However, data in the atlas are
recorded assessments of one or more anthropologists based on local observations at
some time in history. Some of these observations are over 100 years old. As such,
these assessments can vary from current, on-the-ground truth by quiet a bit. For our
purposes, the comparison of model results to anthropological data is primarily of
ordinal value, not interval or ratio value.
!
Farming
Herding
Fig. 6 Results of Household Decision-Making in a Central Burundi
Towards Validating Households
13
6 Conclusions and Future Work
We modeled the weather, parcel type, vegetation, and household subsistence activities in a 1,000 by 1,000 mile area around Lake Victoria in East Africa at the one
square kilometer, one day, and household level. Results from the MASON RiftLand
agent-based model match the types of primary and secondary subsistence activities,
farming and herding, as reported in the anthropological data for the three tested regions. However, we do not yet match the exact percentages of those activities. We
are also continuing our modeling efforts, improving the modeling of the weather,
vegetation, and watering holes behavior as well as re-designing the farming, herding, and household decision-making to support full scale runs of the model in reasonable times and have achieved a nearly 1,000 times improvement. As a result of
our continuing to improve our modeling, we will need to re-validate our model. This
initial modal or ordinal validation in behavioral patterns is a necessary step in the
direction of more extensive validation to be accomplished via additional tests based
on more stringent interval and ratio standards.
Acknowledgements This work was supported by the Center for Social Complexity at George
Mason University and by the Office of Naval Research (ONR) under a Multidisciplinary University Research Initiative (MURI) grant no. N00014–08–1–0921. The authors thank members of the
Mason-HRAF Joint Project on Eastern Africa (MURI Team), especially Jeffrey K. Bassett, Atesmachew B. Hailegiorgis, Joseph Harrison, and Eric Scott for comments and discussions. Carol
Ember, Ian Skoggard, and Teferi Abate of HRAF provided assistance with anthropological data.
The opinions, findings, and conclusions or recommendations expressed in this work are those of
the authors and do not necessarily reflect the views of the sponsors.
References
1. Cioffi-Revilla, C. (2010) MASON RebeLand and Data Aspects of Agent-Based Simulation
Models. In: D. Schmorrow and D. Nicholson (eds.), Advances in Cross-Cultural Decision
Making, pp. 291–301. Orlando, FL: CRC Press-Taylor and Francis
2. Hailegiorgis, A.B., Kennedy, W.G., Catalin Balan, G., Bassett, J.K., Gulden, T. (2010) An
Agent Based Model of Climate Change and Conflict among Pastoralists in East Africa.
In: Proceedings of the International Congress on Environmental Modeling and Software
(IEMSS2010). Ottawa, Ontario, Canada
3. Kennedy, W.G., Bassett, J.K. (2011) Implementing a Fast and Frugal Cognitive Model within
a Computational Social Simulation. In: Proceedings of the Second Annual Conference of the
Computational Social Science Society of the Americas CSSSA-2011, Santa Fe, New Mexico
4. Kennedy, W.G., Gulden, T., Hailegiorgis, A.B., Bassett, J.K., Coletti, M., Catalin Balan, G.,
Clark, M., Cioffi-Revilla, C. (2010) An Agent-Based Model of Conflict in East Africa and the
Effect of the Privatization of Land. In: Proceedings of the Third World Congress on Social
Simulation WCSS-2010, Kassel, Germany
5. Lewis, I.M. (1955) Peoples of the Horn of Africa. London: International African Institute.
6. Little, P.D., McPeak, J., Barrett, C.B., Kristjanson, P. (2008) Challenging Orthodoxies: Understanding Poverty in Pastoral Areas of East Africa. Development and Change, 39(4), 587–
611
14
Kennedy et al.
7. McCabe, J.T. (2004) Cattle Bring Us to Our Enemies. University of Michigan Press: Ann
Arbor: MI
8. Luke S., Cioffi-Revilla, C., Sullivan, K., Catalin Balan, G. (2005) MASON: A Multiagent
Simulation Environment. Simulation 81(7), 517527
9. Luke,
S.
Multiagent
Simulation
and
the
MASON
(2012)
Library.
http://cs.gmu.edu/ eclab/projects/mason/
10. Cioffi-Revilla, C., Crooks, A., De Jong, K., Gulden, T., Kennedy, W. Luke, S., Coletti, M.
(2012) MASON RiftLand: an agent-based model for analyzing conflicts, disasters, and humanitarian crises in east Africa. Working Paper, Mason-Yale Joint Project on Eastern Africa.
Center for Social Complexity, George Mason University, Fairfax, Virginia 22030
11. Murdock, G.P. (1966) Ethnographic Atlas. Ethnology 5(3), 317–345
12. Murdock, G.P. (1967) Ethnographic Atlas: a summary (Codes). International Journal of Cultural and Social Anthropology 4, 154–169
13. Vrieling, A., de Beurs, K.M., Brown, M.E. (2011) Variability of African farming systems
from phenological analysis of NDVI time series. Climatic Change, 109(3-4), 455–477
14. Taylor, B.K. (1962) The West Lacustrine Bantu. London: International African Institute