➤ The University of Virginia Biocomplexity Institute runs three COVID-19-tracking platforms that use AI and machine-learning techniques to project the impact of the pandemic worldwide.
➤ Launched in February, the COVID-19 Surveillance Dashboard features county-level data for the U.S. and state/province-level data for 15 countries.
➤ As the pandemic unfolds, forecasted data is compared to observed data allowing for constant revision and evolution of analytical tools. These tools can help resolve uncertainties.
Christopher Barrett, executive director, Madhav Marathe, division director, and Bryan Lewis, research associate professor, spoke to S&P Global Market Intelligence about the abilities and limitations of their COVID analytical models. Answers are edited for brevity and clarity.
S&P Global Market Intelligence: What kinds of questions are people asking that could be answered with your COVID analytical models?
Bryan Lewis: In the short term, it's, "What's this thing going to look like in a couple of weeks?" They want forecasts, so decision-makers can at least in the short term make some plans. In the long term, it's "What's it going to look like when the flu gets here? How are we going to open up schools because we need something for the kids to do? We need to continue their education, but what's the best way to set things up?"
Madhav Marathe: The questions the state [Virginia] studies are like the questions that the nation studies. If you understand the differences between different counties in the state and the social disparity that accounts for this, like why one part might be lighting up more than the other, the same questions hold at the national scale. It's amazing — this pandemic has brought forward almost every question that an epidemiologist might ever introduce to an analyst.
Could you explain to us how predictive modeling is answering these questions?
Marathe: Prediction is a loaded term, so we have chosen instead to use the word projection because we are trying to tell what might happen under different scenarios. Unlike physical systems where you can pretty much predict exactly the track of where things would go, like where the position of the moon would be, here things are much more complicated. The forecast, or our projection, is endogenous to the system. If we say things are going to get really bad in Virginia, people might change their behavior. Our actions change the outcome of the system.
What are the data sources powering your analytical models?
Lewis: To the degree possible, we use the observed state of the disease and its history, so how many cases there have been, hospitalizations, deaths and more all at the state and national level. Other things including the mobility of the population and a lot of the qualitative things that are more difficult to categorize. One thing we've been trying to do is quantify, or at least set dates for, when the different COVID policies have come into play.
Every time we've done one of these epidemic responses, there's a huge reliance on journalists who capture local info and distill it in a way that we can then use it — there's no database out there that you can just type into and download exact policy provisions across the entire nation.
What are some challenges you are encountering that are difficult to factor in using data?
Lewis: I think one thing being a little under-reported is the morbidity associated with COVID. Any of these people that are suffering from neurological damage and serious lung damage might have life-long health implications. They're not going to be able to fulfill things they wanted to do later in life, and that's sometimes not as easy to account for. You can't just watch the deaths and cases climb on whatever dashboard or tracker headlines you're looking at and think that 3 million people had it and a bunch of people survived. A good chunk of those people who survived are injured and will be injured for the rest of their lives.
Christopher Barrett: One question we are asking is why it is that you would make the decision to do, say, in-person schooling and somebody else identical to you statistically would have the decision to do online schooling. On the one hand, we have so much data, but on the other hand, we don't have anything when it comes to questions like this. And we don't even know where to look for it. This is a huge science challenge.
How can you tell if we're in the first or second wave of the pandemic?
Lewis: I kind of bristle a bit about saying, "there's going to be a first wave, then second wave, then third wave." We're basically a surfer on a continuous wave. Either, we're riding it well and we're going to have a nice easy landing or we're going to get washed up.
At the moment, we're trembling on the edge about to get washed over by it. I've been calling it a surge. I think that this is a continuous process. We can either keep it tamped down, or we can lose control and get dragged away by it. We've never sat on the beach and enjoyed ourselves for a while longer before the next wave comes.