Decision-making in the time of a pandemic

The speed with which the Covid-19 crisis is unfolding is putting the world’s leaders mercilessly to the test. They — and indeed all of us — have to make decisions fast, under great uncertainty, and with great stakes hanging on the outcome. This is above all a practical challenge, but also a theoretical one that economics is supposed to help with. How do we make the best decisions in such conditions?

Coincidentally, this week I reviewed for the FT a new book on precisely that task, Radical Uncertainty by John Kay and Mervyn King — two luminaries of the British economic policy establishment. Their book criticises the standard approach to uncertainty in economics and related disciplines, which requires a comprehensive list of possible outcomes with well-defined numerical probabilities attached. As I point out in the review, this is a necessary critique insofar as probabilistic decision theory can give an entirely unwarranted sense of scientific precision in many real-life situations where the uncertainties are simply too deep. But I disagree that this disqualifies probabilistic reasoning from being taken into account, even in such situations. The question is not whether we should employ probabilistic theory but how we employ it.

The current crisis is a case in point. How the pandemic is going to unfold is unknown and unknowable. The full set of possible outcomes is too complex to enumerate, and it is foolish to think there are precise, objective probabilities for each of those outcomes. (Though Kay and King finished their book before the epidemic broke out, everything they say suggests they would agree with those assessments.) But it would be wrong to think we can make good decisions now without trying to quantify both outcomes or probabilities. On the contrary, only such an attempt can gives us a good handle on how our values and informal judgment should shape what we do.

To see this, take the time to read Imperial College London’s epidemiology modelling which is said to play a big part in the UK government’s decision-making process (also read our news coverage of the report). This is, by all accounts, a solid model. The simple mechanics are easy enough to understand. The model takes survey-based data on where and how much individuals interact and therefore expose each other to contamination, and combines them with medically informed assumptions about the contagiousness and persistence of the virus to mathematically derive how it spreads through the population. Against this benchmark simulation, the authors then check what happens if various interventions such as household quarantining and social distancing reduce contact and hence contagion to various degrees. My colleague Jemima Kelly has a step-by-step breakdown of the analysis.

The numbers that go into this model are at best approximated, at worst made up. Taking Kay and King more seriously than they may have intended, we might conclude that the entire exercise is worse than nothing. But that cannot be right. While we cannot possibly conceive of the full set of possible outcomes in all their richness, we can distil in quantitative terms what matters most about them: the number of contaminated, the number of seriously ill, the number of intensive care beds needed and the number of deaths. No good decision can be taken in the face of this radical uncertainty without quantifying as best we can the need for medical treatment and the consequent likelihood of exceeding the health system’s ability to provide it.

Next, what about probabilities? The Imperial College paper, curiously, does not attempt to describe probabilities. The model is deterministic, using single values both in the parameters and in the outcomes. It assumes, for example, “an incubation period of 5.1 days” and that “50% of household comply with the policy” of household quarantine. Admittedly they try out a few different values for some key parameters (but not for those that describe policy intervention, such as compliance rate). But this is a situation where informed probabilistic reasoning would make us understand the problem better.

We may not know the exact probabilities but it is as reasonable to assume they fall the further you go from the authors’ guesses at incubation periods and compliance rates as it is to make those guesses in the first place. But a probabilistic model would allow us to say which uncertainty the predicted outcome was most sensitive to. Are we more likely to get the prediction wrong because we get the compliance rate wrong? Or because we get the incubation period wrong? Knowing this would help us know where we need to kick the tyres more on the assumptions going into the model.

The UK’s policy towards the epidemic has been controversial, first for seemingly adopting a “mitigation” rather than a “suppression” strategy — simply slowing the pace at which most of the population gets infected versus trying to minimise the number infected — and, second, for taking longer than other countries to put in place the most draconian suppression policies, such as school closures. Formal and probabilistic thinking helps us think about both.

The government has sometimes suggested that new information, specifically about the higher hospitalisation rate among the symptomatic infected, led to the policy change last weekend. But understanding this model’s style of reasoning makes it clear that the earlier mitigation strategy could only be justified with numbers very different from those that were already clear from the Chinese experience (and published in scientific outlets) by late January. As Richard Horton, editor of the Lancet, writes:

“The science has been the same since January . . . it didn’t need this week’s predictions by Imperial College scientists to estimate the impact of the government’s complacent approach. Any numerate school student could make the calculation. With a mortality of 1% among 60% of a population of some 66 million people, the UK could expect almost 400,000 deaths. The huge wave of critically ill patients that would result from this strategy would quickly overwhelm the NHS.”

The Imperial College model produces a result that school closures are less effective in spreading the infection than other social distancing measures. But it partly forces this conclusion by assuming — by its own admission “mechanistically” and pessimistically — that school closures would increase contagion within households or in the community. It does not explore how uncertainty around these parameters may change our prediction of the relative effectiveness of different interventions, nor how these effects might vary in the presence of multiple interventions at the same time. For example, it may be that school closures lead to people having more contact outside of school if no other policies are introduced — but should this not be treated as less likely when strict social distancing and household quarantine measures are also in force?

We do not know how the Covid-19 pandemic will play out, nor how well the policy interventions will work. This may be unknowable even after the fact. As far as that goes, Kay and King’s message is important to heed. But, silly as it may seem to pretend we can accurately quantify outcomes and probabilities, doing so as well as we can still provides a strong basis on which to make and (importantly) to question policy choices, so long as it is not given the final word.