Visualizing the Impact of SARS-CoV-2 Intervention Strategies

 

Overview

The rapid spread of SARS-CoV-2 has led many countries and regions to enact various interventions, such as social distancing, school closures, and border control, in order to mitigate the growth of infection. Understanding the effects of these interventions is particularly important since each strategy comes with its side effects. We wanted to understand the impact of intervention strategies and their combinations on the disease spread. After collecting data at the country and state levels for certain types of interventions, we overlaid them on the disease growth curves, shown above.

Effects of Interventions

As a first step towards understanding the impact of interventions, the visualization above shows the number of confirmed cases over time (on a log scale), for the 30 US states with the most confirmed cases of SARS-CoV-2. We have overlaid trends with various countermeasures taken by the governing entities. We use icons (💼, 🏠, 👨‍👩‍👧‍👦, 🎓, 🍔, 🏬) to depict these interventions; regional interventions are shown in smaller size than interventions that are more encompassing (either at the state level for US States, or at the country level for countries). For example, the icon 🏠 indicates a stay-at-home order or lockdown. We invite the reader to reveal such measures on a per-region basis in the chart by clicking the legend (recommended) or the chart itself. (Select multiple regions by shift-clicking on the legend.) You can switch between the visualizations for countries or US states by switching between the World or USA tabs, and between the visualizations for confirmed cases or deaths by switching between the Cases or Mortality tabs.

To visualize the impact of the interventions, we also plot a projection line for the original trajectory of the trend prior to the last major intervention enacted. The projection extrapolates the growth based on the slope in the last five days prior to the intervention. This projection is based on the simple assumption that the growth rate stays fixed throughout the entire period of time, which is not always a valid assumption for a number of reasons. For example, as the number of infected individuals increases, the growth will likely slow down due to the increasing number of recovered people with immunity. Nevertheless, this serves as one comparison point that we can use to understand the effects of interventions in slowing the infection rates.

Datasets and Procedures

We draw on data regarding COVID-19 cases and deaths from JHU Coronavirus Resource Center as well as the New York Times US dataset.

We manually constructed a dataset for interventions enacted in each US state, drawing from Wikipedia as well as the New York Times. The new dataset we collected is available here (last updated April 16th). This dataset captures information about:

  • Emergency Declarations
  • School, Restaurant, Non-Essential Business Closures
  • Banning of Gatherings
  • Visitor Quarantines and Border Closures
  • Stay-at-home Orders

After several rounds of manual coding procedures for these interventions, we eventually developed a form to force consistent and error-free coding. This form is available here.

We have augmented this dataset to produce an enhanced dataset available here that may be more useful for computational epidemiologists. In particular, we list each specified intervention with a new line for each geographic entity (e.g., US city, county, township, school district, etc.). Where possible, each geographic entity was mapped to a unique Federal Information Processing Standard (FIPS) code, and then merged with available population data from the US Census to identify the total number of people living in the area where the specified intervention was enacted.

We are aware of complementary efforts to collect intervention data at the county level, such as those from Keystone and Stanford; however, these datasets do not cover all counties.

For countries, we drew on Oxford’s Coronavirus Government Response Tracker (last updated April 19th). We mapped their “Restrictions on movement” label to our “Stay at home” label; the “Workplace closing” label was mapped to “Business closures”; and their “Restrictions on gatherings” label was mapped to our “Gatherings banned” label.

FAQs

  • What are the drawbacks of our visualization dashboards?

There is danger in extrapolating too much from limited historical data, especially since many of the case numbers are subject to other confounding variables, such as the amount and availability of tests. We will be keeping the dashboard up-to-date with the latest data to see how these trends unfold.

Another drawback is that our extrapolation (labeled as original trajectory in the visualization) is easy-to-understand but simplistic: other more sophisticated models exist. That said, our intent is not prediction, but rather provide a visual cue to study the differences before and after the intervention.

Finally, we must mention that aggregate patterns and trends often obscure individual datapoints and outliers. Visualizing data on a logarithmic scale, while making it easier to visualize exponential growth, often gives us a false sense of linear behavior.

  • How was the original trajectory computed?

The trajectory was computed by drawing a straight line from the last five days prior to the point of the intervention, and then extending that post the intervention.

  • Why build yet another COVID-19 visualization?

While there are many COVID-19 visualization dashboards, including those that employ helpful log-linear extrapolation to understand the trends in various regions, we haven’t found any dashboards that try to visualize the overlaid visual impact of various intervention measures, apart from anecdotal reports of the curve being flattened thanks to interventions. If there are any visualization dashboards that we should be aware of and can link to, please share them with us at covidvis@berkeley.edu.

  • How can we reproduce the charts above?

Our Jupyter notebooks, processing scripts, and underlying datasets are online on GitHub.

  • How can I contribute?

Please write to us at covidvis@berkeley.edu

Other Acknowledgments

There are many visualizations of COVID-19 growth curves online that we draw on for inspiration. We are fans of visualizations from John Burn-Murdoch, Financial Times, such as this one, as well as the New York Times, such as this, this, this, and this. We drew on data preprocessing scripts from Wade Fagen’s excellent “Flip the script on COVID-19” dashboard.

Our visualization dashboard employs many popular open-source packages, including Altair and Vega-Lite, for visualization, Pandas for data processing, and Jupyter for sharing code and visualizations.

Team

Covidvis is a collaborative effort across computational epidemiology, public health, and visualization researchers at UC Berkeley (EECS, Innovate For Health Program, School of Information, and School of Public Health), University of Illinois (Computer Science), and Georgia Tech (Computational Science and Engineering).

From the visualization side, the team includes Doris Jung-Lin Lee (UC Berkeley School of Information); Stephen Macke (University of Illinois Computer Science and UC Berkeley EECS); Murtaza Ali (UC Berkeley); Ti-Chung Cheng, Tana Wattanawaroon, and Pingjing Yang (University of Illinois Computer Science); and Aditya Parameswaran (UC Berkeley School of Information and EECS).

From the public health and epidemiology side, the team includes Stephanie Eaneff (Data Science Health Innovation Fellow at UC Berkeley and UCSF), Alexander Rodriguez (Georgia Tech Computational Science and Engineering), Anika Tabassum (Virginia Tech Computer Science), as well as B Aditya Prakash (Georgia Tech Computational Science and Engineering). We’ve also benefited enormously from ideas and input from Ziad Obermeyer (UC Berkeley School of Public Health).