Using agent-based models to
investigate high-risk adolescent
partner formation
Amanda D. Latimore, MA PhD
September 26, 2012
© 2008, Johns Hopkins University. All rights reserved.
Background: Concurrency  STIs
Concurrency: the act of engaging in multiple sexual partnerships
that overlap in time
Compared to serial monogamy, given the same total number of
partners, concurrency amplifies STI risk
• Individual concurrency
 increases risk of transmission
• Partner concurrency
 increases risk of acquisition
2%
10%
41%
64%
Percent of people that are connected through
their sexual partnerships
Figure credit: Jim Moody and Martina Morris
Morris & Kretzschmar, 1997; 2009;
Aral, 2010
http://cptoolkit.hivsharespace.net/index.html
© 2008, Johns Hopkins University. All rights reserved.
2
Concurrency  US Racial/ethnic STI disparities
Syphilis
Highest burden in adolescents
3
Centers for Disease Control, 2011
© 2008, Johns Hopkins University. All rights reserved.
Background: Black women prefer
monogamous men
• 2.6 – 3.1 times the odds of concurrency in Black men
(28%) vs. Whites men (13%)
• Adolescent black women express fatalism about the
availability of “good” men
• Monogamous
• Physical attractiveness - proxy for cleanliness,
faithfulness and low STI risk
• As well as a partner with educational and career goals, that
is emotionally supportive, respectful and kind
Then how and why does concurrency occur?
4
Adimora, Schoenbach & Doherty, 2007;
Andrinopoulos, Kerrigan & Ellen, 2006
© 2008, Johns Hopkins University. All rights reserved.
Background: Sex ratio  Concurrency
Proportion of men in the US, 2000
1980
1990
52.0 – 75
51.0 – 51.9
50.0 – 50.9
49.0 – 49.9
47.9 – 48.9
1.0 – 47.6
www.nationalatlas.gov
US Census Bureau
5
© 2008, Johns Hopkins University. All rights reserved.
Poverty, race and sex ratio - US Census
2010
46
42
D
Po
or
bl
ac
k,
M
Po
or
bl
ac
k
hi
te
Po
or
w
G
en
er
al
US
D
Po
or
bl
ac
k,
M
Po
or
bl
ac
k
hi
te
Po
or
w
G
en
er
al
US
70
75
44
80
85
Proportion of men (%)
48
90
95
50
2000
Baltimore City STD clinic-based, adolescents 15-24 years
6
© 2008, Johns Hopkins University. All rights reserved.
Male firearm homicide mortality rate per
100,000 population
Background: Sex ratio driven by violence
and disproportionate incarceration
120
100
1999
80
2000
2001
60
2002
2003
40
2004
2005
20
0
.15-19
White
. 20-24
Black
White
Black
. 25-29
White
Black
. 30-34
White
Black
. 35-39
White
Black
7
Hu, Webster & Baker, 2008
© 2008, Johns Hopkins University. All rights reserved.
Background: Sex ratio driven by birth disparity
Trends in male birth proportion by race and ethnicity
8
Branum, Parker & Schoendorf, 2009
© 2008, Johns Hopkins University. All rights reserved.
Population-level
Individual-level
Social Determinants
PARTNER SCARCITY
Psychosocial Processes
Sex Ratio
theory
&
Demographic
Opportunity
theory
Female TOLERANCE
Male Partner Concurrency
Population-level
Background:
Summary
CONCURRENCY
9
STI Prevalence
Overarching Goal
To understand how context (partner scarcity) interacts with
individual-level partner preferences to form the empirically
observed levels of male concurrency
Population: Sexually active, Black, inner-city adolescents
Exposures:
Proportion of men at the Census tract level
Female tolerance of male concurrency
Outcomes: Prevalence of men engaged in concurrency
Method: Agent-based modeling
10
© 2008, Johns Hopkins University. All rights reserved.
Specific Aims
PART ONE
I. What is agent-based modeling?
II. What is gained by using an agent-based approach?
PART TWO
III. What is the impact of partner scarcity on concurrency relative
to female tolerance and can partner scarcity explain the blackwhite disparity in concurrency?
IV. By what mechanisms can a female preference for monogamy
lead to population-level male concurrency?
11
© 2008, Johns Hopkins University. All rights reserved.
What do a flock of birds, a school of fish and a herd of bison
have in common?
PART ONE
© 2008, Johns Hopkins University. All rights reserved.
I. What is Agent-based modeling?
A multi-agent, stochastic, discrete-event simulation of
individuals/components that are:
• Heuristic
• Adaptive
• Autonomous*
• Interacting
• Interdependent
• Networked
EMERGENT
PHENOMENA
13
Macy, 2002
© 2008, Johns Hopkins University. All rights reserved.
I. Why Agent-based Modeling?
How do you capture using
traditional experimental design?
14
WHO Commission on Social Determinants of Health, 2007
© 2008, Johns Hopkins University. All rights reserved.
I. Why Agent-based Modeling?
15
WHO Commission on Social Determinants of Health, 2007
© 2008, Johns Hopkins University. All rights reserved.
I. ABM utility in complex health systems
Agent-based is a holistic approach
• Model heterogeneous populations
• Understand the dynamic cross-level feedback of the
individual and the micro/mezzo/macro environment
• Study counter-intuitive or emergent phenomenon using a
bottom-up approach
16
© 2008, Johns Hopkins University. All rights reserved.
Why simulate?
Fear-inspired flight and
epidemic dynamics
Epstein, Parker Cummings, Hammond, 2008
Emergency evacuation
Dr. Paul Torrens, Dept of Geographical
Sciences , University of Maryland
Adolescent partner
formation?
17
© 2008, Johns Hopkins University. All rights reserved.
Specific Aims
PART ONE
I. What is agent-based modeling?
II. What is gained by using an agent-based approach?
PART TWO
III. What is the impact of partner scarcity on concurrency relative
to female tolerance and can partner scarcity explain the blackwhite disparity in concurrency?
IV. By what mechanisms can a female preference for monogamy
lead to population-level male concurrency?
18
© 2008, Johns Hopkins University. All rights reserved.
Complexity # 1: Heterogeneity
Advantage: Ability to capture variation meaningful to the outcome
• More granular predictions
• Better interventions for sub-populations and hubs of risk
Considerations: Variation exists in partner formation strategies and risk

• Physical space 
• Social space
• Relationship type and duration

• Within individuals across time and experience
Current application: Variations in partner number, sex ratio, change in
tolerance due to experience
© 2008, Johns Hopkins University. All rights reserved.
19
Complexity # 2: The interaction of individuals
with each other and their environments
Advantage: Individuals do not act in isolation
• Novel levels of influence for interventions
• Better predictions
Considerations: Unrealistic to assume full, instantaneous or completely
random access to resources in the environment

• Network position (CONCURRENCY), size, diffusion
• Movement of individuals through/around built/social environments
• Error or delay in information exchange

Current application: Partial recognition of partner’s partners, dating network
20
© 2008, Johns Hopkins University. All rights reserved.
Complexity # 3: Modeling individual processes
Advantage:
• Closer to understanding high-risk or counter-intuitive behavior
Considerations: Observation may not fully represent underlying process
• Assortative mixing: Do we seek the most attractive or the most similar?
• “Prettier at closing time” phenomenon

• Female acceptance of male concurrency despite preference for monogamy
Current application: Autonomous, locally-constrained decision-making
21
© 2008, Johns Hopkins University. All rights reserved.
System
Dynamics/
Compartmental
Multiple,
Discrete Event
Simulation
Multiple
Microsimulation
Multiple
Individual-level
interaction
Structure &
Adaptive and
autonomous indiv
Network
Analysis
Heterogeneity
One
Feedback
Traditional
Regression
(LDA/MLM)
Explicit
concurrency
Computer Simulation
Systems Science
-
-
-
-
-

-  -

-
 -
-
Dynamic

  
Mixing
matrices
Dynamic

Dynamic

   
Dynamic and
directed
-
 - 
n/a
Static or
discretely
longitudinal
n/a
Static or
dynamic
Differential
equations
Dynamic and
directed
Adaptive,
Queue logic
(Multiple)
relative position
Aggregate
Queue, resource,
entity
Individual
Temporality
Method
Transition
probabilities
Level of analysis
Methodological Approaches
ABM
Multiple
Individual
Adaptive
Heuristic
Markov Family
Individual or
No memory
Aggregate
Independent
of history
Stahl, 2008; Luke & Stamatakis, 2011; Gilbert, 2008; Chattoe-Brown, 2009

-
 -

22
PART TWO
© 2008, Johns Hopkins University. All rights reserved.
Specific Aims
PART ONE
I. What is agent-based modeling?
II. What is gained by using an agent-based approach?
PART TWO
III. What is the impact of partner scarcity on concurrency relative
to female tolerance and can partner scarcity explain the blackwhite disparity in concurrency?
IV. By what mechanisms can a female preference for monogamy
lead to population-level male concurrency?
24
© 2008, Johns Hopkins University. All rights reserved.
Population-level
Individual-level
Social Determinants
PARTNER SCARCITY
AIM 2:
Static model
(One time point)
Psychosocial Processes
Female TOLERANCE
Population-level
Male Partner Concurrency
CONCURRENCY
25
STI Prevalence
Specific Aims
PART ONE
I. What is agent-based modeling?
II. What is gained by using an agent-based approach?
PART TWO
III. What is the impact of partner scarcity on concurrency relative
to female tolerance and can partner scarcity explain the blackwhite disparity in concurrency?
IV. By what mechanisms can a female preference for monogamy
lead to population-level male concurrency?
26
© 2008, Johns Hopkins University. All rights reserved.
Population-level
Individual-level
Social Determinants
PARTNER SCARCITY
?
Psychosocial Processes
AIM 3:
Dynamic
model
(Multiple time
points with
feedback)
Female TOLERANCE
Population-level
Male Partner Concurrency
CONCURRENCY
27
STI Prevalence
=
+
THOUSANDS of lines of code
28
Model Calibration and Validation Data
Description
PRSTD II1
Black, inner-city adolescents (14-19 years)
recruited from Baltimore STI clinics, 20002003
40
30
20
Random sample of HIV negative African
National Survey American state ID holders in 13 central and
eastern North Carolina counties with
of Family
Growth5
registered HIV cases(18-59 years), 19972000
10
US probability
sample4
North Carolina population-based case-control
study (15-44 years), used control data only,
2002
0
National school-based sample of 15-19
years, 1999
Frequency (%)
AddHealth3
50
Population-based sample adolescents (14Sexual network 19), urban San Francisco neighborhood,
study2
2000
35
40
45
50
Proportion of men (%)
60
PRSTD II represented Census tracts (Mean = 47.1, SD = 4.2)
Simulated data (Mean = 47.1, SD = 4.9)
1. Raw data, Lenoir et al., 2003; Bettinger et al., 2004; Matson et al., 2012; 2. Fichtenberg
et al., 2009; 3. Kelly et al., 2003; 4. Doherty et al., 2009a; 5. Doherty et al., 2009b
Lenoir et al, 2006
55
© 2008, Johns Hopkins University. All rights reserved.
29
THE INFLUENCE OF CONTEXT AND COMPOSITION IN THE
RISKY PARTNER SELECTION OF URBAN ADOLESCENTS:
AN AGENT-BASED INVESTIGATION
Amanda D. Latimore, Pamela Matson, Jonathan M. Ellen, Derek
Cummings and David D. Celentano
III. Static ABM
A. What are the relative cross-sectional effects of gender ratio and
female tolerance on population-level concurrency?
B. Can partner scarcity explain the black-white difference we observe in
male concurrency?
30
© 2008, Johns Hopkins University. All rights reserved.
Point prevalence of male concurrency
Hypothesized relative effects
High
Tolerance
Low
Tolerance
Low % men
High % men
Female Tolerance of Male
Concurrency
Partner Scarcity
31
© 2008, Johns Hopkins University. All rights reserved.
Static ABM components
Partner Formation
INPUT
Gender
Ratio
Output
INPUT
Female
Tolerance
Point
prevalence of
male
concurrency
Other network
characteristics
Female degree constrained
32
III. Static model:
Partner formation
Initialization
For an individual in a
population with a
given proportion of
men and female
tolerance…
Seek partners from dating pool
Seek ~ P(4)
Seek 1
Available single ♂?
Yes
Monogamous ♂
Deterministic pairing
Paired
No
Concurrent ♂
Stochastic pairing wp ∝ TOLERANCE
Single
33
© 2008, Johns Hopkins University. All rights reserved.
100
III. Static model: Overall results
0
20
40
60
% Concurrent men
80
96.0
89.4
82.8
76.2
69.5
62.9
56.3
49.7
43.0
36.4
29.8
23.2
16.6
9.93
3.31
60
50
40
Proportion of men in the population (%)
30
KEY FINDINGS:
• Positive association of partner scarcity and female tolerance with concurrency
• Slight advantage of female tolerance?
34
© 2008, Johns Hopkins University. All rights reserved.
III. Results – Absolute change
60
80
100
Change due to sex
ratio when
tolerance is high
0
20
40
“Independent
effect” of partner
scarcity
60
50
40
Proportion of men in the population (%)
+/- 2 standard deviations
30
Tolerance <= 25%
Tolerance >= 75%
0
20
40
60
80
100
Change due to
tolerance when
partners are
scarce
0
20
40
60
80
100
“Independent
effect” of female
tolerance
Female tolerance of male concurrency (%)
+/- 2 SDs
Low Proportion of men (30-40%)
High Proportion of men (50-60%)
35
© 2008, Johns Hopkins University. All rights reserved.
7
7
III (A) Comparing relative effects
Avg RR(low% men)
Avg RR(high tolerance)
(95% confidence range)
2.14
(1.96, 2.37)
2.55
(2.27, 2.99)
6
2.57
(2.17, 3.08)
2.25
(2.07, 2.47)
1.89
(1.68, 2.12)
2
2
3
3
4
4
5
5
6
3.33
(2.70, 4.48)
(95% confidence range)
0
20
40
60
80
100
Female tolerance for male concurrency (%)
(A)
III (B) Black-white disparity
60
50
40
30
Proportion of men in the population (%)
(B)
OR = 2.56 (95% CI: 1.61, 4.07)
Adjusted OR = 3.06 (95% CI: 2.27, 4.13)
Prevalence Ratio = 2.15
(Adimora et al., 2007)
36
© 2008, Johns Hopkins University. All rights reserved.
THE INFLUENCE OF CONTEXT AND COMPOSITION IN THE
RISKY PARTNER SELECTION OF URBAN ADOLESCENTS:
AN AGENT-BASED INVESTIGATION
Amanda D. Latimore, Pamela Matson, Jonathan M. Ellen, Derek
Cummings and David D. Celentano
III. Static ABM
A. What are the relative cross-sectional effects of gender ratio and
female tolerance on population-level concurrency?
• Interacting and difficult to compare
• Similar in magnitude
B. Can partner scarcity explain the black-white difference we observe in
male concurrency?
• Yes. And changes in tolerance just as impactful
© 2008, Johns Hopkins University. All rights reserved.
37
THE NEGATIVE IMPACT OF PARTNER SCARCITY ON
CONCURRENCY IN URBAN ADOLESCENTS:
Guiding interventions using an agent-based model of
high-risk partner formation
Amanda D. Latimore, Derek Cummings, Pamela Matson,
Jonathan M. Ellen, and David D. Celentano
IV. Dynamic ABM
How/why does female tolerance develop?
A. Primary mechanism: RELAXATION hypothesis
 What are the short- and long-term impacts of an effective tolerance
intervention?
B. Alternative mechanism: MISINFORMATION hypothesis
38
© 2008, Johns Hopkins University. All rights reserved.
Dynamic ABM components
RELAXATION hypothesis
Time
Varied INPUT
Gender
Ratio
INPUT
Tolerance =
0 at time0
Output
Prevalence of
male
concurrency
Other network
characteristics
Learning
Female degree
constrained
Relationship
characteristics
39
IV. Dynamic ABM
RELAXATION hypothesis
ijk
jk
Time + 1
Relationship Duration (RD) + 1
Time+ 1
Check partner status
with error
Check RD
Break up
LEARNING
“Prettier at closing time”
Attempts <= Threshold
Attempts + 1
Attempts > Threshold
Tolerance increase
© 2008, Johns Hopkins University. All rights reserved.
40
80
IV. Population-level data
Change in tolerance across time
RELAXATION hypothesis
60
Average age , 50% threshold
40
% Men
0
20
30-40
42.5-47.5
50-60
10
15
20
%
Mean
populations
(SD)
that reach
threshold
16.2 (.44) 100%
18.4 (.97) 100%
21.6 (.66) 19%
25
Age (years)
42.5-47.5% men
30-40% men
50-60% men
mean +/- 2 SDs
41
© 2008, Johns Hopkins University. All rights reserved.
IV. Population-level data
Change in effects across time
RELAXATION hypothesis
42
© 2008, Johns Hopkins University. All rights reserved.
RELAXATION hypothesis
IV. Partner-level data
Survival to and Hazard of male concurrency
When partner scarcity exists:
• Adjusting for tolerance significantly decreases the hazard of male
concurrency in a relationship
© 2008, Johns Hopkins University. All rights reserved.
43
IV. Population-level
Cross-sectional impact of reductions in tolerance
βTolerance = 2.08
(95% range 1.43, 5.35)
Take-home point: 2-to-1 decrease in concurrency for every %point decrease in tolerance, on average, across levels of sex ratio
© 2008, Johns Hopkins University. All rights reserved.
44
IV. Population-level data
Long-term impact of a reduction in tolerance
Populations below 50% tolerance
Simulation time
50
100
150
% men
Intervention
age (step)
30-40
16.1 (50)
N
Intervention: Control
144:156
0
.2
.4
.6
.8
1
0
12
14
16
18
20
22
42.5 – 47.5 18.2 (75)
89:91
50-60
79:28
Age
30-40% men
50-60% men
42.5-47.5% men
Point of intervention
21.6 (116)
45
© 2008, Johns Hopkins University. All rights reserved.
IV. Long-term impact of a reduction in tolerance
42.5 - 47.5% Men
Percentage-point
35 40 45 50 55 60 65 70 75 80 85
Baseline
20
30
40
50
Baseline
20
30
15
12
9
6
3
0
change
Percent
30 - 40% Men
40
50
Intervention time (months)
Key Finding: A one-time reduction in tolerance can have long-term effects on
concurrency
© 2008, Johns Hopkins University. All rights reserved.
46
THE NEGATIVE IMPACT OF PARTNER SCARCITY ON
CONCURRENCY IN URBAN ADOLESCENTS:
Guiding interventions using an agent-based model of
high-risk partner formation
Amanda D. Latimore, Derek Cummings, Pamela Matson,
Jonathan M. Ellen, and David D. Celentano
IV. Dynamic ABM
How/why does female tolerance develop?
A. Primary mechanism: RELAXATION hypothesis
 What are the short- and long-term impacts of an effective tolerance
intervention?
About 2 times the change in concurrency for a one-unit change in
tolerance – maintained over time
47
© 2008, Johns Hopkins University. All rights reserved.
IV (B) Alternate mechanism
MISINFORMATION hypothesis
ijk
jk
Time + 1
Relationship Duration (RD) + 1
Time+ 1
Check partner status
with error
Check RD
Break up
LEARNING
“Prettier at closing time”
X
Attempts <= Threshold
Attempts + 1
Attempts > Threshold
Tolerance increase
© 2008, Johns Hopkins University. All rights reserved.
48
IV (B) Results (47.5% men)
MISINFORMATION hypothesis
Percent of identified concurrent men
0%
21%
42%
63%
100%
Misinformation alone cannot explain the concurrency that we observe
(66% of PRSTD sample)
© 2008, Johns Hopkins University. All rights reserved.
49
IV (B) Results
35% men (7% of PRSTD sample)
MISINFORMATION hypothesis
55% men (2% of PRSTD sample)
50
© 2008, Johns Hopkins University. All rights reserved.
THE NEGATIVE IMPACT OF PARTNER SCARCITY ON
CONCURRENCY IN URBAN ADOLESCENTS:
Guiding interventions using an agent-based model of
high-risk partner formation
Amanda D. Latimore, Derek Cummings, Pamela Matson,
Jonathan M. Ellen, and David D. Celentano
IV. Dynamic ABM
How/why does female tolerance develop?
A. Main mechanism: RELAXATION hypothesis
 What are the short- and long-term impacts of an effective tolerance intervention?
About 2 times the change in concurrency per one-unit change in tolerance,
maintained across time
B. Alternative mechanism: MISINFORMATION hypothesis
Not likely, but a potential point of impact for those exposed to severe partner scarcity
51
© 2008, Johns Hopkins University. All rights reserved.
Summary
© 2008, Johns Hopkins University. All rights reserved.
Population-level
Individual-level
Social Determinants
PARTNER SCARCITY
Psychosocial Processes
Female TOLERANCE
Male Partner Concurrency
Population-level
AIM 2:
Static ABM
1) Interacting and
similar in
magnitude of effect
on concurrency
2) Partner scarcity can
explain the blackwhite disparity in
concurrency
CONCURRENCY
53
STI Prevalence
Population-level
Individual-level
Social Determinants
PARTNER SCARCITY
?
Psychosocial Processes
Female TOLERANCE
Male Partner Concurrency
Population-level
AIM 3:
Dynamic ABM
1) A decrease in
tolerance can have
short- and long-term
impact on
concurrency
2) What we observe is
likely due to both a
relaxation of
preferences and
misinformation
CONCURRENCY
54
STI Prevalence
Limitations
• Need for simplifying assumptions
• From the perspective of the female only
• Stable sexual networks
• Only intra-racial partner formations
• Generalizability
• Limited data available to inform some components
• Serious computing power needed for dynamic models
• Insights limited by current statistical methods
• Almost any result can be “calibrated” into a simulation
55
© 2008, Johns Hopkins University. All rights reserved.
Strengths
• Systems science is the only way to address the feedback
loops present in health systems
• ABMs most appropriate for individual-level complexity
• Model was calibrated and validated with empirical data
• Ability to model concurrency
• Observe generative processes
• Examination of interdependent factors interacting across
levels/time
• Ability to model process of agent learning
56
© 2008, Johns Hopkins University. All rights reserved.
Innovation and public health impact
• Provides support for a common explanation for blackwhite disparity in concurrency and possibly STIs
• Provides insight into an empirically unobserved
mechanism between partner preference and selection
• Novel application of an emerging method
• Application to a growing field that recognizes
concurrency and other network-level STI risks
• Showed the potential effects of an alternative points of impact
• ~*Complexity*~
© 2008, Johns Hopkins University. All rights reserved.
Future Directions – Where do we go from here?
Applications
• Systemic interventions for underlying causes of male scarcity
• Investigate affects of expanding female partner pool
• Groundwork for tolerance measures
ABM extensions
• Investigate individual interventions
• Multidimensional partner characteristics
• Dynamic potential partner networks
• Changes in the accuracy of partner evaluations by intimacy
• Incorporate STI transmission and condom use
• Modify for other contexts – Concurrency is a hot topic!
© 2008, Johns Hopkins University. All rights reserved.
58
Neighborhood, Family, & Individual Factors
Neighborhood
(Joblessness, Urban
disadvantage, etc)
Race
Under-resourced schools
Family Social Support
Age
Gender Ratio
Violence
Academic
competency
Negotiating Power
Incarceration
Pregnancy
Female Tolerance
Marriage
Rates
Norms
Partner Concurrency
Relationship
Duration/Type/
Detachment