Year of growth for virus models
By Dave Orrick
dorrick@pioneerpress.com
They were scary, but sort of comforting. They were wrong, but they were also right. They were relied upon and reviled. And they most definitely grabbed our attention.
We’re talking about the coronavirus statistical models that emerged around a year ago, just as the pandemic was beginning to set upon the world, America and Minnesota. A year later, those models have matured and now account for factors like human behavior — even politics, to a degree — but those behind them acknowledge their imperfections are plenty, and they’re just one tool among many to combat the virus.
Models no longer play the center- stage role they did in the early days, but leaders, including Minnesota Gov. Tim Walz, continue to consult them — and insist that regardless of how close the models foresaw the actual number of deaths or hospital beds occupied, they have saved lives.
‘NO OTHER DATA’
March 2020 was a time when everyone — including political leaders, health officials and certainly the media — were clamoring for information and some sense of the future, and models seemed to provide both. They reduced the horrors of people suffocating alone amid exasperated doctors and nurses to smooth lines on a graph that forebodingly rose but then reassuringly fell.
They sketched narratives that, depending on the model and your perspective — and how well you understood them — ranged from terrifying to not-that-bad for a
PANDEMIC MODELS, 4A
while, followed by a clear conclusion. And they seemed to carry the certainty of mathematics.
Their curves were the target of the slogan “flatten the curve” that permeated much of the ethos supporting stayat- home orders, like those that Walz and governors across the nation unilaterally enacted under emergency powers.
They were also all we had.
The virus that emerged in Wuhan, China, had already mutated to what scientists now suspect was a more infectious strain that took northern Italy by storm and would become the primary variant that spread across the globe — at least for a while. But that bit of information, like so much else about SARS-CoV-2, was essentially unknown at the time.
“We had no other data, no other real-world experience,” Walz said in a recent interview with the Pioneer Press, reflecting on the early days of the pandemic, when testing was in almost as paltry supply as masks and ventilators. The former turned out to be essential in slowing transmission and protecting medical workers; the latter, in hindsight, remained an important tool for treating the sick, but never reached the ubiquitous prominence that it seemed destined for in the early days.
EARLY MODELS
The first model that gained attraction was put forth by Imperial College London. It said 2.2 million Americans would die. Well, actually it projected that, perhaps, 2.2 million Americans could die — if nothing was done. In other words, if we acted like we did every flu season, which is to basically not change anything.
Reactions among political leaders — and their respective followers — ranged from fear to scoffs. The models didn’t predict that America’s political polarization would translate to our pandemic response — but models soon became politicized.
President Donald Trump and others who downplayed the virus frequently highlighted projections by modelers at the Institute for Health Metrics and Evaluation at the University of Washington. That model, which did attempt to take into account how restrictions could lessen the virus’s impact, initially projected U.S. deaths would range from 38,000 to 162,000 during the early months.
With a growing number of models — eventually more than 30 prominent models would be publicly available online — the White House ultimately combined figures from several and last spring settled on a range of 100,000 to 240,000 deaths.
As of Thursday, some 529,000 Americans had died from confirmed cases of COVID-19, according to the New York Times’ tally.
In late March, amid calls of “show us the data,” Walz publicly presented a Minnesotaspecific model developed by the University of Minnesota’s School of Public Health with cooperation from the Minnesota Department of Health.
“There were so many unknowns, but it was nice to try to quantify the risk,” Walz said of that model, which he leaned on heavily in the early days as he ordered businesses, schools and churches to close their doors and most aspects of face-to-face society to cease.
The model, like most at the time, generally envisioned a single curve, with a single peak, at least in the near future. The peak that came in the following months was not as severe. That model projected between 50,000 and 55,000 deaths, a figure that by May would be updated to project 16,000 to 44,000 deaths.
As of Friday, 6,737 Minnesotans have died from COVID19.
UNPREDICTABILITY
Whether any of that means the models were prescient or not isn’t a fair standard, cautioned Curtis Storlie, director of data science at the Mayo Clinic’s Kern Center “It was a darn near impossible task in February to predict this thing,” said Storlie, who helped develop what would become Mayo’s model. He said policymakers were grasping for some sort of guidance, and the basic tenets of epidemiology helped inform them.
The problem with the early models was they were unable to account for the two most important variables in their equations: the virus and us.
The models improved — at least for short-term projections, as the dynamics of the biology came into focus, said Dr. Sean Dowdy, who oversees aspects of quality improvement at Mayo and became involved with attempting to anticipate their COVID response.
“I think if this were all rats in a cage, Curtis could predict this pretty well,” he said of Storlie’s modeling.
But we’re not rats. Cases of the virus rose and fell across the nation and the globe, defying obvious and convenient causes, such as the weather, and with varying degrees of intensity. However, a few trends emerged that statisticians and scientists now understand.
SWINGS IN BEHAVIOR
Many who’ve analyzed the virus’s path now believe in a dynamic relationship between people — especially those who are most likely to catch and spread the virus — and their behavior, based not on some universal rules of human interaction, but on recent history.
Here’s the simplest way to explain it: When cases are low, people relax and the virus spreads. Then people put their guard up, and cases fall. And then they relax.
“As they become more aware and concerned, they change their behavior,” Storlie said in an interview. “And as they get more comfortable, they have people over for dinner. We have settled into a pattern of behavior that we will rinse and repeat in a random but somewhat predictable way. You can’t predict it per se, but you can see and understand the patterns. Implicit in our model is that when things get low, they become more likely to rise.”
Models also began to look at more streams of data, including travel patterns and mask compliance — which itself can be a factor of politics.
The more-enlightened models have shown their worth.
FALL SURGE FORESEEN Some models, including Mayo’s, predicted the early fall surge that rocked the Upper Midwest and portended a nationwide surge that eclipsed all others.
When modelers showed Walz what they saw coming, he said he was skeptical.
“I got a brief in late September that predicted a peak that was unimaginable in Minnesota,” said Walz, noting that for the numbers to come true, the Upper Midwest would have to see rates of spread not seen anywhere else on the planet.
That’s exactly what happened. “They were eerily accurate,” he said.
But the confidence in the level of predictability for infections wanes the farther out a model tries to predict; like the weather, much more than two weeks and it gets more uncertain. Yet, general trends can be foreseen, modelers say, months ahead — like the climate’s changing seasons.
MODELING TODAY
Last week, Mayo decided to make its coronavirus model’s projections available for the public. It incorporates yet another novel variable in the pandemic: vaccinations.
While it does project an increase in cases in coming months, the model doesn’t see a likelihood of a massive surge coming.
However, Storlie acknowledged, the algorithms don’t fully encapsulate the effects of spreading variants, which, depending on the variant, can increase the infectiousness of the virus and increase the likelihood of someone being reinfected.
The emergence of variants is the latest development challenging modelers. Earlier this month, Chris Murray at the University of Washington’s IHME that produces its widely followed model, said new data on the virus’s mutations had prompted him to change his assumptions. The biggest takeaway: He, like a growing number of scientists, is now skeptical that the virus will ever go away and is now trying to envision how immunity-resistant strains might behave, according to an interview he gave to Bloomberg News.
Some, including Dr. Mike Osterholm, director of the University of Minnesota’s Center for Infectious Disease Research and Policy, are clamoring for an end to current relaxations of restrictions, fearing a massive surge in cases, driven primarily by the B.1.1.7. variant first discovered in the United Kingdom.
Storlie said at this point, the Mayo Clinic’s model sees that as an unlikely — but not impossible — scenario. “What we’re seeing the next four months out is a bump,” he said.
That was important for Walz, who consulted with both Murray and Storlie last week as he was preparing to ease restrictions to levels not scene since the pandemic began.
Walz said other data informs him as well now, including positivity rates — made possible by widespread testing — and genetic sequencing, which analyzes about 7 percent of all positive cases, and increasing rates of vaccinations.
“I know there are some who are worried about a big surge,” Walz told the Pioneer Press before announcing his decision but after consulting with the modelers and others, “but I’m just not seeing it right now.”