My favourite lecture ever, and what it says about fighting COVID-19
I'd like to tell you about my favourite lecture of my entire time as a student, on February 20, 1996. It is extraordinarily relevant in today's world, because it described models for epidemics. I remember because it came with a convincing case that much can be learned from this class of models by any student applying models to inform urgent decisions. It's a lecture that I've kept learning from for the last 24 years.
In fact, before I get back to epidemics, that was the first of two really valuable things I learned across that semester:
So, for models - what's missing can really matter. Who'd have thought to model this:
The lecture was given by John Harte at Berkeley, and part of a class on environmental modelling co-taught with Zac Powell. Many may know John for his introductory Consider a Spherical Cow book, and a subset of the lectures eventually became the more advanced Cylindrical Cow book.
I loved it because it felt like drinking math out of a firehose. Apparently, I was thirsty. And in terms of differential equations, these models are archetypal, the maths equivalent of the greatest novels.
I loved it because it felt like drinking math out of a firehose. Apparently, I was thirsty. And in terms of differential equations, these models are archetypal, the maths equivalent of the greatest novels.
Let me first say, don't for a minute dismiss cows as irrelevant: New Zealand's response against COVID-19 may have been sharpened considerably by a painful recent experience, including on contact tracing, as part of billion dollar effort to stamp out the spread of the Mycoplasma bovis incursion in dairy herds across the country.
In fact, before I get back to epidemics, that was the first of two really valuable things I learned across that semester:
- Mathematical models can be powerfully reused and rebuilt, helping new situations seem familiar based on previous experience.
- Models designed to fit exactly the situation you're dealing with are very powerful, and building them helps you see clearly what really matters (or mattered).
And let me just stop here for a moment and say that I'm mainly writing this for experts and science communicators who are suddenly looking at this class of models and trying to converse about them, as well as the art of modelling. For me these epidemic models are like a song that pops into my head every time I think about how to model systems well, or diagnose what's wrong with a model that appears to fit well but remains suspicious. That's because of a lecture 24 years ago. I hope it is a bit helpful allowing experts who are and aren't modellers converse about how models help us with COVID-19, particularly in New Zealand. There are many positive lessons, but it is equally true that we're seeing mistakes and misstatements massively amplified this past weekend in the age of social media in ways that will be worth noting, remembering and teaching.
Now, about the maths of epidemics: the models fit almost every epidemic remarkably well. Yet this is especially true when applied retrospectively, allowing parameters to be reconstructed. Beware however: there can be large uncertainty in early predictions, compounded by delayed or unreliable data.
Worse, there can be confusion when fits are are excellent but for the wrong reason. A good way to find errors is to check that the numbers chosen for parameters and outcomes make sense. Witness the problem when the UK temporarily followed a strategy that assumed it would be useful to develop herd immunity to COVID-19. The number of hospitalised people under such as strategy would quickly be roughly 30 times greater than what hospitals could handle. It also makes me wonder if someone conceptually (or mathematically) misinterpreted the control of spread in China and South Korea as the onset of a resistant population, when it wasn't.
At the moment, I worry about a problem that commonly afflicts New Zealand: the tendency to import models without redesigning them to fit our circumstances. The Imperial College modelling is possibly such a case. It is hugely valuable for the situation the UK and most other western nations suddenly find themselves in. But it describes options that are useful to consider when community transmission has already run out of control – which is not currently the case in New Zealand.
It has become apparent that many people, and apparently entrepreneurs in particular, are looking at the exponential increase we are seeing here, and assuming we have the beginning stages of community transmission. Instead, the case descriptions make it clear that at least 95% of our confirmed cases are linked directly to international travel from origins where exponential growth of infections was well underway. Those arriving passengers show infections averaging 5.5 days after contact, so we should just now be getting to a safer zone where all arriving passengers after Sunday 15 March self isolated. I'm keenly aware of this: I arrived back in the Monday morning cohort, and watched the confirmed cases in the US state I left behind rise from 5 to 30 while I was flying.
So, by all means, let's understand how the Imperial College modelling informs strategies if we pass through alert levels 3 and 4. But more importantly let's be scared by this possibility mainly so we avoid it, because uncontrolled exponential growth of cases is like falling off a high cliff. Instead, we can keep our eyes on path ahead – suppressing the spread – acknowledging that it will be treacherous. It should go without saying that panic (and associated misinterpretation of data) will not help.
We can succeed where others failed, it seems to me. That means social distancing and good hygiene from everyone, following self-isolation rules, and then use contact tracing data to inform models that optimise interventions. That seems the best way stop the spread associated with arriving travellers and any dispersal within New Zealand, particularly across regions. I think the government has been vigilant and proactive, and it is pleasing to say the most important steps are the ones that were taken a week ago.
Models can matter in this, and any battle against time like this. Each time we use them, we learn more about how to make them matter. Models are often requested to simulate – predict the future – and often can't do that. They're often most useful for what they tell us about uncertainty – what we can't predict – and about how things really work so we can help select which of the availability choices decision makers should pursue.
This work is licensed under a Creative Commons Attribution 4.0 International License.
P.S. If you're looking for specific recommendations on COVID-19, this is not the place. This is to encourage thinking about how we use models, and what we look to learn from them. I do suggest you trust (for the most part) the government and scientists who are leading communication on this.
P.P.S.
This work is licensed under a Creative Commons Attribution 4.0 International License.
P.S. If you're looking for specific recommendations on COVID-19, this is not the place. This is to encourage thinking about how we use models, and what we look to learn from them. I do suggest you trust (for the most part) the government and scientists who are leading communication on this.
P.P.S.
So given what I wrote, am I surprised or upset by the announcement to go to level 4 in 48 hours?— Troy Baisden (@TroyBaisden) March 23, 2020
No, models are uncertain.
Models ≠ vigilance.
Precaution is good, and has to be applied in a uniform way people can understand.
And so it shall be.
So, for models - what's missing can really matter. Who'd have thought to model this:
Cases of #COVID19nz at the World Hereford Conference, A&P show, and post conference field trips deserve more recognition. This is what matters. Maybe it provides evidence (rather than shouting) as the driver of a good decision to go to Level 4. #scary https://t.co/PWyMBKk67L— Troy Baisden (@TroyBaisden) March 23, 2020
Comments
Post a Comment