Characterizing the impacts of disease on behavior across scales: Policy, perception, and potential for infection

This study demonstrates that during the COVID-19 pandemic, US social distancing behavior was primarily driven by state-level policies and perceived risk rather than local conditions, suggesting that disease models should account for rational human responses to disease incidence at broader spatial scales.

The Gist

Imagine the spread of a virus like a wildfire. For a long time, scientists focused entirely on the fuel (the virus itself) and the wind (how people move around). But this new research asks a crucial question: What about the firefighters?

In this study, the "firefighters" are regular people deciding whether to stay home, wear a mask, or avoid a party. The researchers wanted to figure out: What actually makes people decide to put out the fire? Do they listen to the local sheriff (county policies), the state governor (state policies), or their own gut feeling about how dangerous things are?

Here is a simple breakdown of what they found, using some everyday analogies.

1. The Big Question: Who is the Boss of Your Behavior?

During the pandemic, people changed their habits. But why?

• The "Local Sheriff" Theory: Maybe people only care about what's happening in their own town. If the local hospital is full, they stay home.
• The "State Governor" Theory: Maybe people only listen to big, broad rules from the state level.
• The "Gut Feeling" Theory: Maybe people just panic or feel safe based on how worried they personally are, regardless of the actual numbers.

2. The Investigation: Looking at the Map

The researchers acted like detectives, looking at data from thousands of US counties between late 2020 and early 2021. They compared three things:

1. The Reality: How many people actually got sick (Objective Risk).
2. The Rules: What laws were in place (Policies).
3. The Fear: How worried people said they were on surveys (Perceived Risk).

They also looked at how counties talk to each other. Do neighbors influence neighbors? Do people who commute to work together influence each other? Or do people who are friends on Facebook influence each other?

3. The Big Surprises

🏛️ The "State Governor" Wins

The study found that people were not mostly reacting to what was happening in their specific little town. Instead, they were reacting to State-level information.

• The Analogy: Imagine you are in a small village. You might not know if the forest fire is 5 miles away in your specific neighborhood, but if the Governor says, "The whole state is on high alert," you pack your bags.
• The Finding: State-wide policies and state-wide worry levels were much better at predicting if people would stay home than local county data. This suggests that during the pandemic, people were looking at the "big picture" (TV news, social media, state announcements) rather than checking their local town square.

🗣️ Fear is a Better Predictor than Facts

Interestingly, how worried people felt was a slightly stronger driver of behavior than the actual number of sick people.

• The Analogy: If you hear a loud crash in the kitchen, you might run to check it out even if it was just a cat knocking over a cup. Your fear of the crash made you move, not the reality of the cat.
• The Finding: When people felt scared, they stayed home. When they felt safe, they went out. The researchers found that "fear" and "facts" usually moved in the same direction, so using the actual numbers (facts) is a good enough shortcut for scientists to predict behavior without needing to survey everyone's feelings.

📱 Friends on Facebook Matter More Than Neighbors

The researchers looked at how counties influence each other. They found that social connections (like Facebook friends) and commuting (people driving to work in other towns) mattered more than just physical proximity.

• The Analogy: If your neighbor gets sick, you might worry a little. But if your best friend who lives in a different city posts on Facebook that their whole family is sick, you might get much more worried and change your behavior.
• The Finding: Information (and misinformation) travels through social networks and commute routes faster and more effectively than it travels through physical borders.

4. What Does This Mean for the Future?

The researchers built a computer model to see what happens if we get these predictions wrong.

• The Good News: Even if we don't know exactly how "scared" people are, we can still predict the spread of the disease pretty well just by looking at the case numbers and assuming people act rationally based on those numbers. We don't need to read everyone's mind to build a good model.
• The Bad News: Because people are reacting to State or National news rather than local conditions, they might be taking precautions at the wrong time.
  • Example: If your local town is safe, but the State is in a panic, you might stay home unnecessarily. Conversely, if your town is exploding with cases but the State is calm, you might go out and spread the virus because you aren't "scared" enough yet.

The Bottom Line

During the pandemic, Americans didn't just look at their own backyards; they looked at the horizon. They listened to the State Governor and their social media feeds more than their local news.

For the future: To stop the next "wildfire," health officials need to realize that people are listening to the "State" voice. Policies need to be coordinated so that a person in one county doesn't hear a different story than their neighbor in the next county. If the message is consistent across the whole state, people will make better decisions, and the fire will burn out faster.

Technical Summary

1. Problem Statement

The integration of human behavior into infectious disease models is critical for accurate epidemiological forecasting, yet it remains challenging due to the difficulty in predicting how individuals respond to novel outbreaks. Key unresolved questions include:

• Drivers of Behavior: To what extent is behavioral change (specifically social distancing) driven by objective disease data (measured risk), subjective risk perception, or government policy?
• Spatial Scales: Do individuals respond to local (county-level) conditions or broader (state/national) information?
• Modeling Assumptions: Can objective disease data serve as a sufficient proxy for subjective risk perception in transmission models?
• Connectivity: How do physical proximity, commuting patterns, and social connectivity influence the spread of behavioral responses?

The study addresses these gaps by analyzing the US COVID-19 pandemic, a period characterized by fragmented policy implementation and heterogeneous information exposure across states and counties.

2. Methodology

The study employs a two-pronged approach: a spatiotemporal regression analysis to identify behavioral drivers and an empirically-parameterized epidemiological model to assess the downstream effects on disease dynamics.

A. Data Sources (Sept 2020 – Jan 2021)

• Response Variable: Social distancing behavior, quantified as the mean number of non-household contacts per day (derived from the Delphi Group's COVID-19 Trends and Impact Survey, $N > 6$ million respondents).
• Predictors:
  • Measured Risk: Weekly COVID-19 case incidence and deaths (NYT data).
  • Perceived Risk: Proportion of respondents reporting they were "somewhat" or "very worried" about infection (Delphi survey).
  • Policy: Binary indicators for specific restrictions (e.g., restaurant closures, worship bans) and a composite stringency index (Oxford Tracker) at both State and County levels.
• Controls: Weather (temp, precipitation), indoor/outdoor activity ratios (Safegraph mobility data), demographics (income, race, education), and population density.

B. Spatial Scales of Information

To determine the "scale of influence," the authors aggregated predictors across five distinct spatial scales:

1. Focal County: Local conditions.
2. State: State-wide conditions.
3. Neighboring Counties: Geographically adjacent counties (homophily).
4. Commuting Counties: Counties connected via daily commuting flows (physical interaction).
5. Socially Connected Counties: Counties linked via Facebook friendship networks (social media information transmission).

C. Statistical Modeling

• Spatiotemporal Regression: Used R-INLA (Integrated Nested Laplace Approximation) for Bayesian inference. The model predicted county-week contact rates while controlling for temporal autocorrelation (RW1) and random state effects.
• Sensitivity Analyses: Tested robustness using mobility data (Safegraph) instead of survey contacts, substituting deaths for cases, and correcting for case underreporting.
• Epidemiological Modeling: Developed a discrete-time Susceptible-Infected-Recovered (SIR) model where the susceptible population was split into "worried" and "unworried" groups. Contact rates were dynamically adjusted based on linear regressions linking risk (measured or perceived) to behavior at different spatial scales.

3. Key Results

A. Drivers of Social Distancing

• Perception vs. Measurement: Both perceived risk and measured risk were significant predictors of increased social distancing. However, perceived risk was the strongest driver.
• Spatial Scale Dominance: State-level variables were significantly more predictive of behavioral changes than county-level variables.
  • State-level perceived risk and case incidence showed the strongest correlation with reduced contacts.
  • County-level risk and policies were largely insignificant predictors in the multivariate model.
• Policy Impact: State-level non-essential business restrictions and manufacturing/outdoor recreation bans correlated with increased distancing. Interestingly, some state-level policies (food/drink, worship) showed counter-intuitive associations, potentially due to enforcement heterogeneity or public fatigue.

B. Inter-County Influence (Intermediate Scales)

• Linked County Effects: Conditions in linked counties (neighboring, commuting, or socially connected) were better predictors of focal county behavior than the focal county's own conditions.
• Connectivity Pathways:
  • Socially Connected Counties: Perceived risk in socially linked counties was a stronger predictor of behavior than in geographically adjacent counties, suggesting social media drives risk perception.
  • Commuting Counties: Policy restrictions in commuting counties were strong predictors of behavior in the focal county.
  • Neighboring Counties: Measured risk (cases) was a stronger predictor here than perceived risk.

C. Epidemiological Dynamics

• Interchangeability of Risk: When simulating disease transmission, models using measured risk (cases) produced epidemic dynamics (peak size and timing) nearly indistinguishable from those using perceived risk.
• Spatial Scale in Models: While state-level data explained more behavioral variance in the regression, the choice of spatial scale (county vs. state) in the transmission model did not drastically alter the final epidemic outcomes, provided the behavioral response function was consistent.
• Rationality Assumption: The results support the assumption in many disease models that individuals act "rationally" based on objective data, as objective case counts can effectively substitute for subjective worry in predicting transmission outcomes.

4. Key Contributions

1. Empirical Validation of Behavioral Drivers: The study provides robust evidence that during the US COVID-19 pandemic, behavior was driven more by state-level information and subjective perception than by local county conditions.
2. Identification of Information Pathways: It demonstrates that non-geographic connections (commuting and social media) are as critical, if not more so, than physical proximity in shaping behavioral responses to epidemics.
3. Modeling Guidance: It validates the use of objective case data as a proxy for subjective risk in coupled disease-behavior models, suggesting that complex, hard-to-collect perception data may not be strictly necessary for accurate transmission forecasting.
4. Policy Implications: The findings suggest that localized surveillance and county-level mandates may be less effective than coordinated state-level policies, as the public responds to broader information signals.

5. Significance and Future Directions

• Public Health Strategy: The results argue for coordinated public health messaging and policy at the state or regional level rather than fragmented county-level approaches. Since people respond to broader scales, localized interventions may be undermined if neighboring or socially connected areas have different conditions.
• Modeling Best Practices: Future disease models should incorporate mesoscale (state/regional) information sources rather than assuming purely local decision-making. However, models must also account for spatial heterogeneity in risk tolerance, as assuming uniform responses across all counties leads to inaccurate projections.
• Information Ecology: The study highlights the role of social media and commuting networks in spreading risk perception, suggesting that misinformation or accurate risk communication can propagate rapidly through these non-geographic channels.

Conclusion: The US population modified social distancing behavior primarily in response to state-level case incidence and perceived risk, rather than local conditions. While subjective perception is the primary driver of behavior, objective disease data serves as a reliable substitute for modeling purposes, provided that the spatial scale of information (broader vs. local) is correctly parameterized.