Thread by @docbaty

Something that has been bothering me throughout the course of this pandemic, which has crystallised for me over the last 48 hours: the models we're using for our public health advice are simplistic.

Let me explain...

I'll start by noting that my first degree was a B.Sc majoring in physical applied mathematics and applied statistics. Essentially, to model the real world using equations of the deterministic and stochastic fashion.

(If you ever thought I came across as a bit of a nerd, now you know why)

The models that are being published by Doherty Institute and Burnett and individual epidemiologists all fall into a bucket of models I would describe as simple, linear, and deterministic.

Nature is none of those things. It is complex (and complicated), non-linear, and carries a high degree of randomness (stochastic elements). All systems involving humans are also all of these things.

So where does that send us down the wrong path?

Linear relationships behave in a predictable way. A big change in input leads to an equally big change in output. Small change in input leads to a small change in output.

So if, say, the number of ICU patients goes up a little bit, we might expect quality of care to go down a little bit.

We might also expect case numbers to go up in a steady fashion; until they start to go down.

Simple models are useful in that they support simple narratives. Complex models might give a better forecast, but they're harder to explain. And this is where we're getting into trouble...

We're leaning on simplistic models to drive complex social, public health, and economic decisions, instead of using a suitability complex (and nuanced) model. And so we see a focus on narrative, which then leads to duelling interpretations.

(To put this into context, the modelling we're seeing publicly is the kind of thing I'd expect to see from a 1st or 2nd year Maths student. By 3rd year, I'd be expecting something that includes a probability distribution; confidence intervals; non-linear relationships.)

I see none of why I expect: Reff rates that change over time; the dispersion over time of vaccines; the impact on ICU of tightening capacity; the economic impact on households of an extended lockdown. These are all modelled as static variables...

... all with linear relationships. And so all of those models are grossly inadequate to the task of making state or national decisions.

I’d love to wrong about this, by the way. If you’ve seen models that look like they’re more what I’m advocating for, please let me know!