Ross Hammond, Washington University in St Louis
Abstract:
In this presentation, Professor Hammond will give an overview of recent innovations in agent-based modeling (ABM) applied as a practical tool to inform policy interventions for complex societal problems. He will draw on his current and recent work in population health including work on obesity prevention, health disparities, tobacco control, and pandemic containment.
Bio:
Ross A. Hammond, Ph.D. is the Betty Bofinger Brown Distinguished Professor in Public Health at Washington University in St. Louis, as well as Director of the Center on Social Dynamics & Policy and Senior Fellow in Economics at The Brookings Institution. Hammond’s research applies complex systems modeling tools to generate new insights into the social dynamics that drive many difficult policy problems, and to identify potential leverage points or windows for intervention. Recent work has focused on health disparities, chronic disease prevention, the food system, implementation science, and pandemic containment. He has published extensively in high impact journals such as Lancet, JAMA, Nature Medicine, Science, and PNAS, and his work has been featured in The Atlantic Monthly, The Economist, and New Scientist. Professor Hammond has served in a number of formal policy advisory roles including: as a Special Government Employee at the FDA, on the NIH Advisory Council on Minority Health and Health Disparities, on the Food and Nutrition Board of the National Academies, on the Lancet Commission on Obesity, and as chair of a working group on Dietary Guidelines at the USDA. Hammond holds an additional academic appointment at the Santa Fe Institute, and has taught ABM at Harvard School of Public Health, University of Michigan School of Public Health, and the NIH.
Summary:
Real-world systems are very complex
Each phenomenon is driven by many interacting causes
Interventions are not made in isolation; need a systems approach to understand them
Long-run impact of policies differs significantly from short-run effects; need to think about adaptation
Motivates development of flexible, dynamic agent-based models to capture these complex dynamics
Use-case: Tobacco
More tobacco retailers than McDonalds
Policy question:How do we reduce their density?
Experiment: Tobacco town
Use synthetic populations extracted from census data
Added details about city structure, distribution of tobacco retailer, their products and prices
Watch people move across city, past tobacco retailers
Can change distribution of retailers, prices and products
Evaluates impacts of interventions
Found that combining multiple interventions are far more effective than any one
Use-case: Pandemic containment
TRACE: Testing Responses Through Agent-based Computational Epidemiology
Focused on evaluating policies, not forecasting
Mask
Vaccine distribution
…
Evaluated 12k policy combinations
TRACE evaluates probabilistic samples, so establishes a confidence interval on the impacts of policy choices
300k scenarios
Enables policymakers to design robust policies
Agent-based model:
Allows lots of heterogeneity and dynamic behavior
Realistic networks of contacts:
Home, work, social, etc.
Heterogeneous susceptibility, resistivity, etc.
Different vaccine distribution policies, responses to vaccines
TRACE policy dashboard
Visualizes model results
Enables exploration of possible policy alternatives
Use-case: Obesity
Increasingly common, chronic disease
Moving towards a “whole of community” approach for obesity prevention
Can work with people in their daily lives, rather than in a hospital or similarly limited setting
COMPACT: Childhood Obesity Modeling for Prevention And Community Transformation: https://www.brookings.edu/compact-childhood-obesity-modeling-for-prevention-and-community-transformation/
Interventions more stable when mediated by communities
Stakeholder-driven Community Diffusion
Must bring together the right people, have the right conversation to reach the right groups
Need to develop new measurement tools and models to characterize communities and interventions
Intervention Model:
Represent a community as a graph of groups
Members have a “knowledge” property
We choose steering committees, where members exchange “knowledge”, then return to their groups and spread it.
After intervention community is more informed and connected
Observations
Formal models help to
Capture thinking and evaluate its implications
Preserve learnings across studies
Challenges:
Data limitations
Dialogue with stakeholders
Other applications:
Organizational decision-making
Crime
Education
Corruption
Literacy
…
How does one come up with the dynamics used in models?
Consultation with experts to leverage decades on collected knowledge
Use focused-domain causal inference studies to identify individual drivers
Can use ABMs to explore the dynamics scenarios under which a given randomized controlled trial will produce a given result
Models provided by experts will likely be compact and unify many smaller phenomena into easy to manage/understandable representations
Consultation with individual community members or surveys (e.g. via language models) will elicit large numbers of scenario-specific models that don’t generalize
How do we validate ABMs and gain trust in their fidelity?
Lots of consultation with stakeholders to ensure that key dynamics are represented
Complex models are very sensitive to initial conditions, which are unmeasurable to any real prediction in practice, so it is infeasible to use them to make accurate forecasts
This complexity/stochasticity is a feature, not a bug since it mirrors the real stochasticity of the real world and presents it to stakeholders
A viable validation approach is to compare the predictions of different models and ensure they largely agree