Supplementary online appendices
“Climate Shocks and Migration: An Agent-Based Modeling Approach”
10-3-2015
A1. The Agent-Based Modeling Approach
Agent-based modeling, the core analytical approach for this study, is a technique for the
investigation of the short- and long-term, multi-dimensional, micro- and macro-level
consequences of climate change. First developed in the computer sciences and now commonly
used in geography and other disciplines, agent-based models are relatively new in the
demographic sciences, with some significant exceptions (An and Liu 2010; Aparacio et al. 2011;
Billari and Prskawetz 2003; Billari et al. 2006; Bruch and Mare 2006; Kniveton et al. 2011;
Macy and Willer 2002). Broadly, ABMs are simulations of a population of autonomous,
heterogeneous agents that interact with each other and their environment according to a set of
prescribed rules. The dynamic actions of agents at the micro-level and response to the behaviors
of other agents and characteristics of their environment result in regularities or emergent patterns
at the macro-level.
One key feature of ABMs is that they can create a laboratory for experimental studies of
social phenomena for which researchers could never undertake true experiments with real
humans. As a computational technique, they allow researchers to simulate population processes
over a period of time, then re-simulate these processes with a specific change in one process
while holding constant everything else about the model specification. This creates an
experimental situation with a true control and a true treatment simulation on the same
population. In our case, we use an ABM to examine outcomes of different experimental
scenarios of weather events, which makes the model particularly valuable for research on the
potential short- and long-term consequences of climate change when it is inappropriate, if not
impossible, to wait for these consequences to play out over time. We combine this approach with
experimental testing of theoretical perspectives.
ABMs also allow for the direct incorporation of feedbacks which are fundamental to the
dynamism of human and ecological systems. These feedbacks are of two types. One involves
endogenous relations among key variables: for example, the risk of migration depends on
household assets, which in turn depend on prior migration and remittance behavior. The other
type of feedback involves interaction among agents, or, the influences that neighbors have upon
each other. With feedbacks, ABMs provide the ability to analyze the dynamics of an
interconnected system over time, making it possible to find emergent trajectories that would not
be possible with statistical regression.
A final benefit of ABMs is that a researcher can test the effects of theoretical scenarios,
which can help isolate what is driving the results of a particular experiment. For example,
removing a parameter or changing a particular rule may fundamentally alter the outcomes of the
model.
Figure A1. Flowchart showing overall structure of ABM
Agent Initialize Module
At Year t = 0: Initialize model, create agents and populate attributes values
For each household agent randomly assign a land split trigger (1-5)
Year t x
For each household j in
households list
Household Module
Allocate parent’s assets
TO
If the last parent dies, following will
give to child who cares parent:
1. agricultural land
2. productive and consumer assets
3. dwelling unit (currently no)
No
If there is living child in the village;
randomly choose one
No
If there is living migrant, pick one
Individual
plot module
Individual
person module
Check if trigger condition is TRUE
If so, split land C or C+1 shares
Find the child
If any child is living in the parent’s
household, randomly choose one
TO
Yes
1.4
Household Split occurs:
1. Create a new household
for the subfamily
2. Split 15% household
assets
3. Randomly set a land split
trigger value (1-5)
Find the oldest subfamily;
its probability E3= 0.9
Last parent
dies?
1.3
Random
number
No
E3: risk of
HH splitting?
Yes
2+ subfamilies
Number of
subfamilies
Calculate the E3
only one subfamily
No
1.5
No
Last adult
child?
E10: Calculate wealth
Find the closest relatives
Yes
No
No
Last j ?
Yes
Year
<6
No
Yes
Yes
No
E3=0.33
E3=0.66
E3=0.1
E3=0
Update kinship network for next t
1.6
1.7
Has
children
Stop
Yes
Last t ?
No
Figure A2. Flowchart showing initialization process of ABM
Agent Initialization Module
Agent definition files
Individual.java
At Year t=0
Create village, households, individuals,
parcels and cells agents using agent definition
files; all attributes are set as null initially
1. Household ID
2. Person identifier
3. Age
4. Sex
5. Marriage
6. Migration years
7. Residents status
8. Relationship to Head
9. Spouse ID
10. Father ID
11. Mother ID
Household.java
1. Household ID
2. Assets
3. Roster
4. # of children
5. # of persons
6. # of kin ties
7. Owned Parcel list
8. Managed Parcel list
9. a list of died Person
10. a list of leave couple
11. a list of subfamilies
12. # of remitters
13. land split trigger (1-5)
Village.java
1. Village ID
2. Household list
3. a measure of social
network cohesion
4. the proportion of
migrants who are
remitting
5. the proportion of
certain land type
Parcel.java
1. Parcel ID
2. Land type
3. A list of LulcCell
4. HHID_1
5. HHID_2
6. Yield
7. Size
LulcCell.java
1. Land use type
2. Age
3. Parcel ID
4. HHID_1
5. HHID_2
6. Suitability Score
7. Distance to Road
8. Distance to river
9. Patch ID*
Populate village, household and individuals
agents attributes values by survey date
Populate parcel and cell agents attributes
values by GIS data
Various raster files used to initialize
LulcCell agent and parcel agent attributes
Create social network using socio-matrix
The socio-matrix includes two basic socio
matrices: Parents and Spouses
Household module
How many lulc cell attributes decide how many raster files needed
Figure A3. Flowchart showing step-by-step process of individual behaviors in ABM
Individual Person Module
household j from
Household Module
For each living
individual i in j.roster
Fertility
rates
Is a Female?
Survival
rates
No
Yes
Birth ?
Random
number
Dies?
Random
number
Is a
migrant?
Set status migrant; next
year, they have same
pattern as migrants but
have to return together
New
household
15%
No
E1: risk of
marriage?
Random
number
Is single?
Yes
No
E4: risk
of away?
Yes
Is single?
E1:risk of
marriage?
Yes
No
Stay
subfamily
15%
No
Yes
1. Create a new person by “like
marrying like”
2. Add this couple to j.leaveCouple;
3. Remove i from j.roster
Random
number
No
1. Create a subfamily for
this couple
2. Update subfamily list
Increase migration years by 1
Yes
No
No
1. Create a new household for
this couple in the same village
2. Randomly set a land split
trigger value (1-5)
E6: risk
of return?
Yes
Random
number
Yes
1. Add I to j.roster this year
2. Update the socio-matrix
no
No
Urban area
40%
Yes
Yes
1.Remove i from j.roster
2. Add i to j.diedPerson list
Removed from
E2: PNR ?
roster, but stay in the
Stay/away
calculation of kinship Rural area
30%
Random
number
No
E5: risk of
remittance?
Yes
Go to E3
Update # of remitter: In later version, create
subroutine to deal with different remitters
Go to E3 left
Last i ?
Yes
Return to Household Module
3
Figure A4. Flowchart showing step-by-step process of household behaviors for each plot of land owned in ABM
Individual Plot Module
Migrants remittances
from Individual
Module
Effects on wealth
Household j from
Household Module
Household wealth
Random
number
Land becomes as
fallow or rented
No
E7: risk of engaging
in agriculture?
Effects on agricultural
intensification
Yes
No
Fore each plot p in
j.plots list
Random
number
Return to
Household Module
Yes
Last p ?
Update
p.landtype = other
other
E8: probability
of one of three land
uses
rice
Update p.landtype = rice
p.yield = yield
E9:
calculate the Yield
Calculate
intensification
Use fertilizer,
pesticide and/or
herbicide
Intensifying
effects on yield
4
A2. Description of changes to village crop yields due to climate shocks in Experiment A.
As shown in Figure 4, both floods and droughts have immediate effects on village crop yields.
Rice yields respond in a clear pattern where weather shock influence decrease rice yield.
However, after the weather shocks are over and the climate returns to a normal-normal pattern
each year, starting in year 17, rice yields from the flood and drought scenarios do not return to
parity with the reference scenario. This is initially curious. In addition, cassava and sugar yields
also display patterns that might not be initially expected. We explain these unexpected patterns
here.
First, considering rice yields, the lack of full recovery is because some households
switched land use in response to the dramatic conditions, converting rice paddies to cassava and
sugar, and did not switch back, and also because many households became poorer and could not
fully fertilize their crops.
With cassava, we find a similar pattern of immediately decreasing yields during the
simulated periods of drought, flood, and variability, and rebounds in production after the climate
returns to normal in year 17. However, as shown in Figure 4, drought conditions decrease total
village cassava yield to a much greater extent than flood conditions. Notably, in all three of the
climate scenarios, total village cassava yields after year 17 are higher than in the reference
scenario. This is a result of farmers switching what they grow in response to the weather; when
conditions return to normal, the newly converted land is productive with the new crops and there
is no motivation to switch back.
Total village sugar yields behave differently than rice and cassava. As shown in Figure 4,
they actually increase during all three of the climate scenarios. After the flood or drought
subsides, total sugar yields experience a second and more dramatic increase after year 17. This is
similar to cassava production after the climate normalizes and is caused by households switching
from rice, and cassava, into sugar production.