Abstract
In response to the pandemic development of the novel coronavirus (SARS-CoV-2), governments worldwide have implemented strategies of suppression by non-pharmaceutical interventions (NPIs). Such NPIs include social distancing, school closures, limiting international travel and complete lockdown. Worldwide the NPIs enforced to limit the spread of COVID-19 are now being lifted. Understanding how the risk increases when NPIs are lifted is important for decision making. Treating NPIs equally across countries and regions limits the possibility for modelling differences in epidemic response, as the response to the NPIs influences can vary between regions and this can affect the epidemic outcome, so do the strength and speed of lifting these. Our solution to this is to measure mobility changes from mobile phone data and their impacts on the basic reproductive number. We model the epidemic in all US states to compare the difference in outcome if NPIs are lifted or retained. We show that keeping NPIs just a few weeks longer has a substantial impact on the epidemic outcome.
Introduction
As the initial wave of the COVID-19 pandemic progressed in early 2020, policymakers faced the critical challenge of determining when to ease "lockdown" measures. Traditional epidemiological models often treat non-pharmaceutical interventions (NPIs) as binary, uniform events. However, public adherence and regional responses to mandates vary significantly. This study addresses this heterogeneity by using mobility data as a continuous proxy for social distancing behavior, allowing for a more nuanced understanding of how "opening up" influences the basic reproductive number ($R$) and subsequent mortality rates across the United States.
Methodology and Data Integration
The researchers utilized a hierarchical Bayesian framework to model the epidemic across all 50 US states simultaneously. The model's primary innovation is the use of relative mobility changes from mobile phone data to estimate $R_t$ (the reproductive number at time $t$).
Two primary datasets informed the model:
- Google COVID-19 Community Mobility Reports: Anonymized trends across six sectors: retail and recreation, grocery and pharmacy, transit stations, workplace, residential, and parks.
- Johns Hopkins University (JHU) COVID-19 Data: Daily reported deaths and confirmed cases, used to back-calculate infection rates through Bayesian posterior distributions.
As shown in the figure above, mobility dropped sharply across most sectors in mid-March. While the model initially considered six sectors, the "Parks" sector was eventually excluded due to erratic, weather-dependent behavior that introduced excessive noise. The remaining five sectors provided a robust proxy for social distancing, with a Pearson correlation of approximately 0.7 between mobility changes and $R$ during the initial lockdown phase.
Regional Heterogeneity and Counterfactual Analysis
A core component of the study is the counterfactual analysis, comparing actual reopening outcomes against a "Continued Lockdown" scenario. In this scenario, mobility levels were maintained at the extreme lows observed immediately following NPI implementation.
The results highlight a stark nonlinearity in viral spread. Small increases in mobility can push $R$ back above 1.0, leading to exponential growth. The timing of reopening proved critical; waiting just a few additional weeks allows the "starting point" of infections to drop low enough that the subsequent increase in mobility results in a significantly flatter curve.
Quantitative differences between these scenarios were profound:
- New York: Under actual lifting conditions, modeled deaths were 14.2% of the previous peak. Continued NPIs would have reduced this to 4.6%.
- North Carolina: This state experienced a massive resurgence, with deaths reaching 185.7% of the previous peak. Under continued NPIs, deaths were modeled to remain at 42.9%.
- Louisiana and Michigan: Both states saw deaths at roughly 43% of their peaks; continued NPIs would have brought those figures down to approximately 7% and 10%, respectively.
Conclusion and Limitations
The study demonstrates that the duration of NPIs is a primary determinant of epidemic outcomes. Maintaining restrictions for even a short additional period can drastically reduce the scale of a resurgence.
However, the authors note several limitations. The model is sensitive to reporting noise in daily death data and does not explicitly account for variables such as Personal Protective Equipment (PPE) usage or changes in clinical treatment protocols. Furthermore, as mobility shifts became more gradual during the reopening phase, the relationship between movement and $R$ became noisier, suggesting that future models should integrate more granular data on the quality of social distancing rather than just the quantity of movement.
Summary updated