But based on research just published in Chaos: An Interdisciplinary Journal of Nonlinear Science, Rebecca Morrison, an assistant professor of computer science at the University of Colorado Boulder, thinks that most of the models in use are not nearly as accurate as they should be.
Working with Americo Cunha, an assistant professor at Rio de Janeiro State University in Brazil, Morrison developed a mathematical technique dubbed an "embedded-discrepancy operator" that she believes is capable of improving the results. She and an international team of colleagues are currently adapting the methodology to the challenge of predictions related to the novel coronavirus.
"It's an incredibly complex problem, and I don't think the embedded-discrepancy approach I'm working on is a cure-all," Morrison acknowledges. "But I hope this can be another tool that modelers and decision-makers will have moving forward. And right now, we can use all the help we can get."
The original research conducted by Morrison and Cunha focused on a different malady: zika, a mosquito-borne virus that shares commonalities with dengue and yellow fever.
According to Morrison, while trying to calibrate some of their model's parameters, Cunha and some of his students discovered that "there was a huge discrepancy between the model and the actual data from the Brazilian Ministry of Health about how many people were actually infected," even after trying to match up the assorted figures — a process made difficult by those embedded discrepancies.
"That's where I came in," Morrison continues. "My work says, 'Well, we have these kinds of pieces that make up our model...but we need to get into the guts of the model and modify the equations to try to come up with better consistency between the model and the data."
Upon the completion of those efforts, Morrison and Cunha needed to test the embedded-discrepancy operator — and fortunately, they had the information to do so. "This was not a true prediction case; we were doing it after the fact. So it was a good test case," she says, adding that because their results were much more accurate when using the mathematical procedure, she's confident "they'll be applicable in a predictive modeling scenario. And that's hopefully what we'll get to soon, as we're working on the COVID-19 example."
Of course, Morrison won't simply be able to plug in the embedded-discrepancy operator, since COVID-19 has so many variables. "It's pretty well understood at this point that most places are under-tested," she notes. "So we don't have a lot of the reliable information we need: how we should count people, are they susceptible, have they already been exposed, are they demonstrating antibodies and, if so, can we count them as recovered? So we'll have to think about that from the modeling side to figure out how we can trust our data when we know it's been under-sampled."
Nonetheless, Morrison is hopeful that she'll be able to apply the embedded-discrepancy operator to COVID-19 in a matter of weeks. Once these kinks are smoothed, we may have a much better idea about whether the predictive models are wrong — and how wrong they are.