Chaitanya Kaligotla | Beedie School of Business, SFU
Jason Ho | Beedie School of Business, SFU
Charles B. Weinberg | Sauder School of Business, UBC
Title: Inference and Validation in Agent-Based Models: Applications in Public Health, Epidemics, and Marketing
Date: Friday, March 20th, 2026
Time: 1:30PM (PDT)
Location: ASB 10900
Abstract:
Agent-based models (ABMs) provide a flexible framework for studying complex social systems in which aggregate outcomes emerge from the interactions of heterogeneous individuals following explicit behavioral rules. Their ability to represent micro-level heterogeneity such as social networks, preferences, and institutional constraints, makes them particularly useful when analytical tractability requires overly simplifying assumptions. However, once ABMs are brought into contact with empirical data, they raise fundamental statistical challenges: simulators are stochastic, likelihoods are typically intractable, and the mapping from micro-level parameters to macro-level outcomes is often highly nonlinear.
This seminar presents three empirically grounded ABMs that illustrate different approaches to model construction, calibration, and validation. We begin with CommunityRx, a model of non-pharmaceutical public health information diffusion that augments a pragmatic clinical trial with an agent-based simulator. Active learning methods are used to explore the model’s parameter space and develop validated in silico experiments. We then present CityCOVID, a large-scale agent-based model of the Chicago population that uses empirically validated synthetic populations and endogenous contact networks to study how activity patterns, school reopening, and protective behaviors influence COVID-19 transmission and hospitalization outcomes. Finally, we introduce WatchMoviesTogether, a model of the theatrical movie market in which group formation, shared consumption, word-of-mouth, and advertising jointly shape box-office dynamics. Unlike prior literature, which focuses on individual decision-making, this model treats co-viewing groups as the fundamental unit of consumption, motivated by the empirical observation that fewer than 13% of moviegoing incidences involve a single viewer. Social networks are constructed from egocentric survey data on moviegoers' co-viewing relationships, and group formation emerges endogenously from agents' interactions on this network.
Across these applications, we discuss practical approaches for linking agent-based simulators to empirical data, including parameter exploration, calibration, and validation strategies for stochastic models with intractable likelihoods. Together, the examples illustrate how ABMs can serve not only as exploratory simulations but also as empirically grounded in silico laboratories for studying complex social systems and informing policy and decision making.