Researchers at MICOM are studying the heterogeneous effects of non-pharmaceutical interventions (NPIs) on epidemic dynamics across counties in the U.S. Because surveillance data only reflect policies that were actually implemented, the team generates counterfactual time series for policies that were not implemented during the COVID-19 pandemic, using a calibrated agent-based model.
The framework produces realistic “what-if” scenarios across more than 150 U.S. counties by combining real-world demographic, mobility, epidemiological, and policy data. It incorporates underreporting correction, time-varying mobility networks, and network-level implementation of NPIs — giving public health researchers a resource for studying the potential effects of alternative intervention strategies. Beyond this use case, the team packages their generated counterfactuals into EpiCF-Bench, a benchmark that fills a key gap in time-series causal inference research by providing realistic counterfactual data for evaluating causal inference methods in both single- and multi-policy settings.
The work has been published at Association for Computing Machinery’s Special Interest Group on Knowledge Discovery and Data Mining (ACM SIGKDD) 2026. To support open science, the team has released the full simulation framework, calibration pipeline, data generation process, and datasets as open source.
Simulation framework: github.com/complex-ai-lab/epi-diff-abm
Causal inference benchmark: github.com/complex-ai-lab/epi-cf-benchmark
More information can be found in their open-access research article: dl.acm.org/doi/abs/10.1145/3770855.3817522