Whitepaper · June 2026
Predicting Public Health Compliance: A Behavioral Hindcast Validation Study
Can agent-based simulation predict population-level behaviors during a major crisis? In this validation pilot, Tiresias simulated county-level health compliance and vaccine uptake during the COVID-19 pandemic using pre-2020 psychographic signals. The results achieved an equal-county Pearson correlation of r = 0.86 (and population-weighted correlation of up to r = 0.93) against ground-truth public outcomes.
In the behavioral sciences, predicting real-world compliance during a crisis is notoriously difficult. Traditional demographic models (using age, gender, and income) typically explain less than 15% of behavioral variance, yielding weak correlation coefficients of r = 0.30 to 0.40. By modeling the psychological structures underlying decision-making, Tiresias achieved an equal-county Pearson correlation of r = 0.86 (and population-weighted correlation of up to r = 0.93) on vaccine adoption rates and compliance—meaning our simulation explained a substantial portion of the behavioral variance across key county test cohorts.
1. Research Design & Data Ingestion
To ensure a strict validation check, the simulation was designed around a retrospective holdout protocol: models were evaluated against county-level behavioral outcomes from the pandemic era.
We built our representation of county populations using three primary pre-pandemic datasets:
- Priors (Rentfrow et al. 2013): Academic Big Five personality norms to establish state-level distributions.
- Demographics (US Census ACS 2019): Census county-level 5-year averages (mapping age, gender, education, and median household income).
- Local Ideology (OpenFEC): Anonymized campaign donation statistics to capture political leanings and local civic engagement vectors.
2. High-Level Simulation Infrastructure
Rather than treating populations as a single homogenous block, the Tiresias simulation pipeline synthesizes representative agent swarms using a bottom-up, agent-based model (ABM):
The county population is segmented into distinct demographic cells (e.g., college-educated females aged 35–44 with high income brackets) using joint probability tables derived from ACS Census data.
Big Five personality distributions (Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism) are imputed for each cell, using Rentfrow state-norms and local FEC indicators as priors.
Pre-pandemic online text patterns are analyzed using Large Language Models to extract Big Five trait shifts. These are combined with the demographic priors using Bayesian Fusion to adjust cell trait means and variances.
Individual agents are instantiated with these psychographic traits. Agents form small-world contact networks and run a daily loop where compliance behaviors spread via social contagion and personal susceptibility.
In the simulation, an agent's likelihood to adopt a health behavior is driven by Conscientiousness (adhering to rules and guidelines), Agreeableness (desire to protect the community), and Neuroticism (responsiveness to threat), moderated by the density of compliance behaviors in their local contact network.
3. Ground Truth Validation Targets
To validate our simulation's predictions, we compared the aggregated agent outcomes against independent ground truth outcomes from two public datasets covering 3,126 US counties:
- CDC County-Level Vaccination Rates: Socrata API vaccination percentiles (fully completed primary series populations).
- The New York Times Mask Usage Survey: Geolocated survey data representing counties' mask compliance rates (aggregating \"Always\" and \"Frequently\" responses).
4. Predictive Performance Outcomes
We evaluated our agent swarms across a diverse set of baseline counties representing over 15 million citizens: Los Angeles County, CA; Cook County, IL; Harris County, TX; Maricopa County, AZ; and San Diego County, CA.
| Evaluation Model / Split | Ground Truth Dataset | Pearson Correlation (r) | Notes |
|---|---|---|---|
| Demographic-Only Baseline | CDC & NYT County Data | r ≈ 0.35 | Legacy prediction models relying strictly on demographics. |
| Equal-County Pearson r | CDC & NYT County Data | r = 0.86 | Tiresias retrospective evaluation across 3,126 US counties. |
| Population-Weighted Pearson r | CDC & NYT County Data | r = 0.93 | Accounts for lower statistical reliability in small-population counties. |
These metrics confirm that Tiresias agent swarms successfully predicted county-level public health behaviors. By modeling the underlying psychometric dynamics, the system achieves a degree of generalizability that legacy demographic models simply cannot replicate.
5. High-Level Replication Guide
Researchers can replicate this study using publicly available components:
- Ingest 2019 ACS census variables for county demographics, constructing demographic cells across age, gender, education, and income.
- Map cell Big Five personality averages using state-level norms (Rentfrow et al. 2013) adjusted by local campaign donation density (OpenFEC).
- Extract psychometric signals from pre-2020 geolocated text footprints using standard LLM API classification prompts.
- Initialize an agent population matching these cell frequencies, structured as a small-world network (Watts-Strogatz topology, clustering coefficient = 0.65).
- Iterate behavior transmission time steps and aggregate county outcomes. Compare results directly against CDC Vaccination and NYT Mask datasets using standard correlation packages.
For questions regarding replication details, data schemas, or custom psychometric weightings, please contact the Tiresias engineering and science briefing team.