In our new paper published in Nature Human Behavior “The unequal effects of the health-economy trade-off during the COVID-19 pandemic”, we address some highly debated issues concerning pandemic management by building a high resolution and data-driven epidemic-economic model through an interdisciplinary collaboration that involved expertise ranging from economics and epidemiology to network and computational sciences.
An interdisciplinary collaboration
At the onset of the Covid-19 pandemic, the economists in our team built an agent-based model to analyze the pandemic's economic impacts. The model started from industry-specific estimates of supply and demand shocks, and was able to forecast the propagation of these shocks ahead of time better than most competitors. However, the initial economic model only included basic epidemiological elements, and we quickly acknowledged the need to add more realistic epidemiological features to understand the tradeoffs between health and the economy. For this reason, in the summer of 2020 I (Marco Pangallo) reached out to Alberto Aleta, who was working on a model incorporating individual mobility data to realistically calibrate contact patterns, thereby offering a detailed representation of the virus's spread. The data enabled him and his team to convincingly simulate the effects of interventions like testing, contact tracing, quarantine measures, and to analyze the role of super-spreader events in specific settings, including workplaces and food-related establishments.
This spurred a multi-year interdisciplinary collaboration that has grown to 12 authors with expertise ranging from epidemiology and economics to computational data science. The common interest in agent-based modeling and mechanistic simulations immediately led us to build a very detailed joint epidemic-economic agent-based model of the New York metropolitan area, grounded on highly granular data on the mobility patterns and socioeconomic characteristics of the population and on the industrial structure of the New York economy. This was not an easy step. The epidemic and economic models had to be put in communication: The epidemic code was written in C, the economic code was in Python, and we managed to couple the models through a specific API. Beside the technical issues, the conceptual challenge was the introduction of the behavioral feedback mechanism that we dubbed fear of infection: as the epidemic starts spreading, individuals get worried and cut back on their consumption of goods and services that require physical contact with others, and this in turn changes the mobility patterns and so epidemic spreading.
Non-pharmaceutical interventions (NPIs) and behavioral feedback
Fear of infection was also at the center of heated debates around lockdowns and other NPIs. According to some, NPIs were needed to contain the pandemic and limit economic impacts. If NPIs were not imposed, individuals would have avoided risky consumption anyway because of fear of infection; in turn, this would have brought up a worse economic recession and many more infections and fatalities compared to the case in which NPIs were enacted early. According to others, NPIs were only damaging the economy, because at-risk, old, individuals would have avoided infections due to fear, while little-risk, young individuals could have kept consuming as normal. Spontaneous behavior change due to fear of infection could have led to the best epidemic and economic outcomes.
To us, the only way to scientifically address this problem was to build a quantitative model based on real-world data that could simulate both disease spreading and economic behavior at a fine-grained level, for instance distinguishing the mobility patterns and consumption behavior of young and old individuals. Whether fear of infection leads to a worse economic recession than NPIs is a quantitative question. Whether fear of infection successfully protects at-risk, old individuals is a quantitative question.
What we found
We conducted counterfactual analyses where we manipulated two variables: the extent of economic shutdowns and the intensity of infection fear. Our hypothesis was that more stringent Non-Pharmaceutical Interventions (NPIs) would yield the most favorable outcomes in both health and economy by closing non-essential businesses. Alternatively, if the behavioral modifications triggered by infection fear alone were effective, then amplifying such behavioral feedback should result in improved epidemic and economic indicators, like reduced Covid-19 mortality and lower unemployment rates. However, our simulations revealed that both mechanisms lead to similar outcomes: stricter economic restrictions and heightened fear of infection both led to increased unemployment and decreased fatalities. Both behavioral and mandated interventions show very heterogeneous effects across population strata and geographical areas. For low-income individuals, the consequences of heightened infection fear or more rigorous economic shutdowns were amplified, resulting in a greater number of lives saved but also a significant increase in job losses. In contrast, for high-income individuals these differences are much smaller. This divide is reflected across geographical differences within New York City: high fear or stringent shutdowns trigger a surge in unemployment in lower-income boroughs like Queens and the Bronx, compared to the more affluent Manhattan. Conversely, when fear was mitigated or restrictions were relaxed, the distribution of unemployment rates across the city became more uniform.
The model was pushed to also investigate behavioral heterogeneities across individuals. For instance, we studied the case in which younger individuals exhibited low levels of infection fear, whereas older individuals—more susceptible to severe Covid-19 outcomes—had high fear of the disease. Advocates of leveraging behavior as a mechanism to control the pandemic consider this scenario ideal. Yet, the improvement in outcomes was only mild: older individuals tended to decrease their social interactions belatedly, often after the virus had already reached them or other members of their household.
Our model also made it possible to explore the effect of other pandemic control policies. For instance, we found that closing manufacturing and construction results in a substantial spike in unemployment with only a marginal decrease in fatalities. Additionally, implementing protective measures late in high fear-of-infection scenarios leads to a dual blow of increased deaths and unemployment.
These results underscore a crucial distinction between behavioral changes and NPIs: while behavioral changes are a result of self-organization, NPIs can be implemented as soon as needed for highest effectiveness.
Our model shows that the availability of detailed data allows building agent-based models to study mitigation strategies and behavioral feedback during a pandemic. ABMs thus provide a versatile and dynamic tool that can be indispensable in strategizing pandemic response, but we must clearly stress that there is no model that fits all pathogens. The conclusions derived in our study might not stand with different infection fatality rate, pathogen transmission mechanisms, or specific intervention strategies not considered in our study. Specific strategies and scenarios should therefore be investigated and discussed with policy makers, continuously updating the model with real-time data. The development of such interdisciplinary approaches that encourages collaboration between epidemiologists, economists, public health experts, and data scientists, would surely lead to a more holistic approach to pandemic management.