The predictions which come out of models of epidemics are often highly sensitive to minor changes in assumptions, so can rightly be accused of being wildly wrong when measured against the eventual outcome. "Improve the model" is a common plea. Of course, the most recent model any team publishes is already a presumed improvement on the previous one. Modelers know they have to improve by being more sensitive as to how they combine the ingredients to bake the perfect cake. For example, early conceptual 1927 modelling was based on assumptions known to be wrong, namely that every person was equally susceptible, and equally capable of spreading infection. Not so, but modelers have to start somewhere in order to improve things further along.
A German team under Prof Streeck argues that workplaces don't spread the coronavirus as much as play spaces: singing in a choir produces an aerosol and spray cloud, dancing together in a room or bar, or nightclub apres-ski, also creates infective clouds, as would any confined space where lots of people are in close contact breathing heavily. Although public health guidance has been coy on this matter, orgies are probably best avoided.
On that theme, there are settings in which you are likely to get a big dose, a large viral load, and others where the globules will be few and far between. As one caustic virologist observed, getting out into the open air is a good defense against respiratory transmitted infections: avoiding infection is a walk in the park.
In that vein, we should start categorizing risks, and alter normal interactions in that light. Make interactions calm, not heavy breathing ones. Improve ventilation where interactions have to take place. Use barriers and find a way to make people-traffic flow more easily.
Now that we have such interest in what different models predict, we might overlook a key question: when and how do we know the final outcome? Some of the important data comes in weeks, such that 4 or 5 weeks can reveal a lot about the progress of the infection. That is not to be sneezed at.
However, the Spanish Flu was being re-evaluated decades later, and although we might be much faster this time around, there will be delayed consequence which may take a decade to unravel, such as the cost to health of a long interruption in normal health monitoring and treatment. It might be worth looking at dental health. Will autopsies a decade or two from now show fewer sound teeth as a consequence of not being able to visit the dentist? Will there be an increase in obesity related deaths due to lack of exercise and over-eating in confinement?
On a broader front, models which look at the spread of an infection are important, but do not model the management of an epidemic. Management involves understanding how food delivery systems can be altered to favour vulnerable populations, or changed to establish effective national rationing; understanding how reducing public transport may cause greater overcrowding as essential workers try to get to work all at the same time on fewer trains; it involves ensuring that there are plentiful stocks of essential equipment such as masks, protective clothing and respirators, and ways of producing new equipment if that proves necessary; launching nation-wide tracking of cases; hardship payments for those unable to work in face-to-face occupations; and the organisation of gradual quarantine lifting. It must consider whether hardship payments are better than declaring a moratorium on rents and taxes during lockdown, and other novel policies.
At a Royal Society lecture in 1990 I was charmed by a diffident presentation given by Prof Dietrich Dorner who described not his success in modelling the future, but the difficulties subjects had when they tried to manage fairly simple models.
His first example was extremely simple. Subjects had to turn a control dial so as to get a small target to move from the top of the screen to a line drawn horizontally across the middle. Clearly, the solution was to turn the knob clockwise so that the target sank down to the mid-line. This proved difficult. Dorner had arranged the system so the target only moved after a delay. Most subjects found this very confusing. They kept turning the dial to get the target down, only to find that after a period of no response the target suddenly shot down past the desired mid-line to the bottom of the screen. Irritated and confused, subjects then twisted the dial anti-clockwise, thus making it shoot back to the top of the screen. It took many corrections and much time to get the target onto the desired mid-line. A minority of subjects made just one cautious movement and then waited to see what happened. Such subjects were able to place the target onto the mid-line quickly and with very few moves.
This was a beautiful illustration of the key feature of executive power, that even if what you command has a real effect on a complex system in the real world, it is usually a delayed effect. Oil tankers take time to turn around.
His next simulation was a factory that made a limited number of products which could be produced in different numbers at different prices. Most subjects jumped in with changes aimed at making the factory more profitable, usually with poor results. Some became obsessed with minor indicators, a premature focus which blinded them to larger problems.
DÖRNER, D. (1990): The logic of failure. In: BROADBENT, D.E., BADDELEY, A. & REASON, J.T. (Eds.): Human Factors in Hazardous Situations. Proceedings of a Royal Society Discussion Meeting, Philosophical Transaction of the Royal Society, London. B 327, 463-473, Oxford: Clarendon Press.
Prof Dorner kindly sent me a simulation of running a country which I used with executives and government policy teams. Results were generally poor and often disastrous. The main problem was people pushing a view of how they wanted the country to function, rather than how it actually functioned. I was working with a government in waiting, and I think they were chastened by their misplaced optimism. With any luck, it was a good warning that formulating policies must always be based on a good understanding of how systems actually work.
Perhaps it is time to see the predictive models as simply part of policy making, informative but not definitive, and to consider all the difficulties which beset governing. More of that in my next post.
Meanwhile: How is Europe doing? Europe has been monitoring weekly death rates since 2015.
Until recently, Covid-19 deaths were less than flu outbreaks, but have now surpassed them in much of Europe. Belgium, France, Italy, Netherlands, Spain, Switzerland and UK are all above previous flu peaks. This is alarming, since Influenza A and B strains are wide-spread, while Covid-19 is mostly active in specific foci of infection, for the moment at least.
It looks as if most of Europe is doing less well than Asia, though Austria, Denmark, Estonia, Finland, Germany, Greece, Hungary, Ireland, Luxemburg, and Malta have done well. Some of these differences will depend on what each plane brought in to which airport, but if these differences remain after a month or so, it would be good to look at the policy decisions they made, to see if one can detect a pattern in a stream of decisions, namely: a strategy.