The PM has indicated that the Australian #COVID19 modelling will be published during the week. A thread (part 2; note part 1 with my disclaimers) https://twitter.com/peripatetical/status/1243684739602501633?s=20

There are going to be two instinctive reactions to the studies, both wrong.
(These are the reactions I have to every model I see, even with some experience in looking at them)

The first reaction is "this assumption/parameter is wrong, how can we believe anything the model says?"

It's appropriate to look at the various parameters and make sure they are in the ballpark of what they should be. All models need to make simplifying assumptions and estimates, and the relevant question is how much they make a difference.

For example, a common assumption is that every individual has an equal chance of meeting another - patently false.

This can be important if spread in households is very different to spread between households. But in a population, this simplifying assumption doesn't usually affect the model output much.

Parameter values can be very difficult to estimate. There is no model that will tell you exactly how effective physical distancing measures are. Have they have reduced transmissions by 25%, 50% or 75%?

Even looking at the data doesn't help much - look at the credible intervals in this report on EU lockdowns - the interventions look like they have worked, but hard to estimate by how much https://www.imperial.ac.uk/mrc-global-infectious-disease-analysis/covid-19/report-13-europe-npi-impact/

The second reaction is "OK, we need 4372 ventilators by the 6th of June" or "this will be over on the 25th of September". This is putting too much faith in the models. They are designed to explore possibilities, and aren't predictions.

There are too many uncertainties to make accurate predictions, and even small changes in parameter values can change things considerably. Precise numbers give an illusion of accuracy.

So, how should we use mathematical models of transmission?

The first is what an unmitigated scenario looks like. Clearly this wouldn't happen now - all countries have taken considerable action to flatten the curve, but it gives us an idea of why we're taking the extreme measures we have.

The second is that all models consistently show that the better the control, the longer this will go on for. This reinforces the need for a sustainable suppression strategy.

The third, and probably most important, is the "what if?" questions. What if we only did isolation and quarantine? What if there was more asymptomatic transmission than we thought? What if more/less people went to ICU than we thought?

It's important not to confuse model inputs with model outputs. If we assume children transmit as much as adults, then closing schools will have an effect. But that's because of the original assumption, which may not be true. https://www.imperial.ac.uk/mrc-global-infectious-disease-analysis/covid-19/report-9-impact-of-npis-on-covid-19/

The fourth is to work out how much we need to do to get this under control. This is actually an easy formula.

If, on average, 1 case causes 2.5 secondary cases (ie R0=2.5), then at this early phase we need to prevent more than 60% of infections to get the epidemic under control (because 2.5 x 0.4 = 1).

Even if we could get R < 1, this doesn't mean everything would be fine. We would have to maintain these interventions to keep it there. And its only an average - there will continue to be transmissions and outbreaks and sick patients in hospital.

If we don't quite get to the magic R0<1, then these models then give us an idea of what might happen, and how quickly, and how much harder we'd have to go to get things under control.

In terms of preparation, we don't need predictive models. We have data. There are already a range of possible futures, from Italy/New York, to Germany/South Korea, to Singapore/Hong Kong.

We still need to prepare for the worst while hoping for the best. A best-case scenario in public health terms would be a perception that we overreacted. But that would mean avoiding the Italy/NY scenario.

Models are only one guide on what to do. They plot out possibilities based on assumptions. But there are many other considerations. How sustainable is what we're doing? How do we get people to physically distance? How do we know where we are and if what we're doing is working?

What are other countries doing well or not so well? How can we mitigate the impacts - the lost jobs, disrupted lives and other harms? How quickly can we tool up hospitals for a possible surge?

All these questions may be informed by models, but the answers are elsewhere.