Networked Systems as Witnesses: Internet Traffic and Compliance with Government Stay-at-Home Measures

The coronavirus disease 2019 pandemic (Covid-19) has radically altered daily life. Starting in the early weeks of 2020 the virus rapidly spread around the world, triggering a number of large-scale interventions to mitigate the spread of the disease starting with the full lockdown of Wuhan, China, where the virus was first reported. Countries around the world quickly followed with measures that range from school closures to large-scale lockdowns.

In the US, the first confirmed cases were reported on January 20, 2020, in Snohomish County, Washington State. By March 23rd, every state had reported cases of the virus. By early April of 2020, several state, city, and county governments had started to adopt similar social-distancing interventions, such as shelter-at-home and temporary closing of non-essential businesses and schools, although with variable levels of enforcement.

While non-pharmaceutical interventions (NPI) such as these are considered the most effective ways to limit the spread of infectious diseases, their use is generally controversial given the political, ethical, and socioeconomic issues they raise. Part of the challenge is the non-obvious link between the level of compliance with such measures and their effectiveness. Studies suggest that vaccines alone will not end the pandemic, and a recently published database on the state of global vaccination reveals that, as of 2024, only ~70% of the world population and 32.7% of people in low-income countries has received some dose of an approved vaccine, making the need to gauge the effectiveness of NPIs all the more pertinent.

This research effort was motivated by a simple observation: as people comply with social distancing measures, one would expect to see a significant change in networked systems’ demand as parents and children work remotely. Thus, the goal of our work was to explore the potential of users’ demand for networked services as a proxy for the social distancing behavior of communities, offering a new approach to evaluate the effectiveness of non-pharmaceutical interventions, with the long-term goal of providing public health authorities with the basis for an evidence-based analysis of their effectiveness.

This research project leveraged the vantage point of one of the largest worldwide CDNs, which receives trillions of content requests per day, to understand changes in demand before and after official lockdowns or stay-at-home policies take effect at different locales. To understand the relationship between these public health measures and network traffic, we relied on natural experiments, comparing compliance and network traffic changes across states and countries, over time.

Our work made the following contributions:

  • We demonstrated the strong relationship between CDN demand and user mobility as a proxy of social distancing.
  • Relying on CDN demand as a new proxy of social distancing, we showed the strong correlation between demand and the growth rate of COVID-19 cases in a selected set of counties in the US.
  • We investigated the effectiveness of school closures as an NPI by exploring the relation between inter-region mobility, CDN demand, and the incidence of COVID-19 infections before and after university campus closures.
  • Finally, we extended prior work studying the effectiveness of mask mandates by using CDN demand to show the combined values of social distancing and mask-mandates.

Sample of results

The following are a small sample of our results and key findings. The full set can be found in the associated publications.

We first explored the association between CDN demand and user mobility leveraging data on demand users impose on the CDN and tbe Google’s Community Mobility Reports of mobile users’ behavior. The following figures show examples of the similarity between the trends of percentage difference of mobility and percentage difference of demand for the month of April 2020 – Fulton County (Georgia) and Montgomery (Pennsylvania), and May 2020 – Fairfax County (Virgina), and Suffolk County (New York). Even in the case of Montgomery County, with 0.39 correlation, the alignment of both curves is apparent with demand and mobility following closely opposite trends.

Following Badr et al., we used daily new cases to derive growth rate ratio (GR) defined as the logarithmic rate of change (number of newly reported cases) over the previous 3 days relative to the logarithmic rate of change over the previous week. We evaluated if and how well demand affects the rate of new infections in different counties in the US, calculating the distance correlations of GR and CDN demand. The following figures are four examples of the opposing trends of GR of Infection Cases and CDN demand for April and May 2020 for Wayne County (Michigan), Passaic County (New Jersey), Miami-Dade County (Florida) and Middlesex (New Jersey).

Despite some prior work on the effectiveness of school closures in reducing the spread of viral respiratory illnesses, the value of this NPI is still disputed. As part of our analysis, we used university campus closures in large college towns in the US, to understand the applicability of CDN demand in evaluating the effectiveness of interventions that result in inter-region mobility. The closure of university campuses results in a different type of mobility which, we expect will have a very different relationship with demand. As campuses close and people potentially relocate outside the region, one would expect to see a correlated drop in demand (rather than an increase, as when people stay home). Following prior work on the analysis of the potential benefits of school closure we used daily COVID-19 incidence per 100,000 residents in each county. We focused on days before and after the second closing of uni- versity campuses, around the Thanksgiving holiday of November 26th, 2020.

The next four figures illustrate examples of school/non-school networks demands and confirmed cases at the University of Illinois, Urbana- Champaign (Champaign County), Cornell University, Ithaca (Tomkins County), University of Michigan (Washtenaw County) and Ohio University (Athens County). Beginning dates on x-ticks correspond to Mondays. We noticed that the counties where these universities are located showed a sharp increase in confirmed cases before and during the closing of their respective campuses.

We leveraged CDN demand as a proxy of social distancing to extend an often-cited study by Van Dyke et al. on the positive impact of mask-wearing. The authors investigated trends in COVID-19 incidence in Kansas before and after a governor’s executive order requiring masks or other face coverings in public spaces came into effect. Among the key limitations of their study, the authors pointed to the absence of social distancing information, a clear confounding factor. We built on their work using CDN demand to address this limitation, and explore the combined effects of social distancing and mask mandates.

The following figures show trends in 7-day average of new COVID-19 cases per 100,000 population among mask-mandated/non-mask-mandated and high-demand/low-demand counties in Kansas. The plots includes cases before and after the governor’s executive order requiring masks which took effect on July 3, 2020. The top row shows non-mask mandated counties with low and high demand on the CDN. The bottom row shows mask-mandated counties with low and high demand on the CDN.

The analysis suggests that, while mask mandates can help reduce the rate of infection (top-right figure), the combined interventions (bottom-right) are significantly more effective at reducing the spread.



This work was supported in part by the U.S. National Science Foundation under Award CNS-2027922. Any opinions, findings, and conclusions or recommendations expressed are those of the authors and do not necessarily reflect the views of any funding agencies.