For a rare time since the start of the pandemic, the National Institute for Excellence in Health and Social Services (INESSS) was unable to establish forecasts on the evolution of the pandemic because of “too unstable data”. On the other hand, the National Institute of Public Health of Quebec (INSPQ) was able to make its own predictions. Why such a disparity in the science of predictions? Decryption.
The crystal balls of INESSS and INSPQ are “extremely different,” explains Mike Bénigéri, director of the office of clinical-administrative data for INESSS.
His team uses so-called “pure statistics” calculations. Based on the data of new hospitalizations, new cases and the positivity rate of the last few days, they can predict this same data for the following days. “This model is good only in the short term, notes Mr. Bénigéri. […] In times of change, it cannot be used either, because it is based on what has happened statistically in recent weeks. Any attempt to predict the future therefore becomes invalid if a curfew, a new vaccine or a new variant quickly changes the situation.
“On the other hand, in a period of stability, it will be easy to use and quite good at predicting bed occupancy”, specifies Mr. Bénigéri.
That’s the difference between the two models. The objective of the statistical model is to anticipate the increase in hospitalizations in order to support the health network. The model used by the INSPQ rather aims to describe the “dynamics” or the “trends” of the pandemic in more or less pessimistic scenarios.
A complexity that rhymes with uncertainty
The model on which the INSPQ is based was developed by the team of Dr. Marc Brisson, from the Research Group in Mathematical Modeling and Health Economics Related to Infectious Diseases. Its approach is called “mechanistic” and is based more on assumptions than on raw data. His calculations answer questions such as: if vaccination increases, if Omicron is three times more contagious, if schools are closed, what will happen?
The research group therefore also calibrates its forecasts with hospitalization and death data, but refines its data processing with other “key elements”.
His team manages to circumvent the uncertainties around the number of declared cases by assuming the statistical bias. They thus start from the premise that only a third of the cases of COVID-19 have been officially registered in Quebec. They manage to estimate this difference by comparing the number of cases recorded with the seroprevalence data from Héma-Québec. Blood donors constitute a representative group of the general population, which makes it possible to verify the reliability of official statistics.
This way of doing things is much more complex, because it uses this “biased data” as feedback to guarantee the reliability of the predictions. “We look at whether the cases that we project into our model are reflected in the observed trend in cases. And it has to make sense,” says Marc Brisson.
This is how he and his colleagues cast a wide net in order to arrive at more general prognoses. “It tells us if it’s going up or down. »
However, as with its “statistical” equivalent, “mechanistic” modeling has evolved over the course of the pandemic to include other influences, such as vaccines, variants or health measures.
The arrival of Omicron and the saturation of screening systems have thus destabilized the two models. For INESSS, this instability proved to be too great for the organization to continue to provide advice. For the INSPQ, hospitalization data has been able to take over so that we can continue to predict the future, with, in return, less reliability in the advanced scenarios.
Thus, the reopening of schools is the very last problem to be solved for Marc Brisson, because the absence of data on the circulation of the virus among the youngest distorts the hospitalization data. “Among the oldest, there is a proportion of older people who are hospitalized. So hospitalizations and deaths represent trends in adults. But there are so few children who are hospitalized that it is important to have other sources of data,” he says.
It should be noted that most states in the world use these two models, “statistical” and “mechanistic”, together to predict the next trends of the virus.
This text is taken from our newsletter “Coronavirus mail” of January 17, 2022. To subscribe, click here.