[LINK] The government's coronavirus modelling

[Roger Clarke]

On 10/4/20 11:21 am, Bernard Robertson-Dunn wrote:
> Here's a much more useful description (in the current circumstances) of
> what's going on than any of Dr Jansson type models
> 
> The one COVID-19 number to watch
> https://www.abc.net.au/news/2020-04-10/coronavirus-data-australia-growth-factor-covid-19/12132478

It's a great reductio ad absurdum.

Pity it's dated 10 April, rather than the morning of 1 April.

Like all simplistic single-data-item representations of complex 
real-world systems, it's so useless it's laughable.

They don't declare a model, just a formula:
"Calculating the daily growth factor is as simple as taking today’s new 
reported cases and dividing it by yesterday’s new cases"

And the model that's implicit in that formula is limited to daily 
snapshots of one small element of the Susceptible, Exposed, Infected, 
Recovered (SEIR) notion.  (They appear to mean by 'cases' something like 
'that subset of Infecteds that have been noticed and recorded').

And, as per my email of 08:57, even the SEIR notion is insufficiently 
complex to provide a basis for useful modelling;  so the Elvery et al. 
tiny sub-set of SEIR is just plain silly.

And that's only the *modelling* aspect of their proposal.  Even if the 
model made some kind of sense, the data it depends on is utter nonsense.

Editing a para. from my email of Fri, 3 Apr 2020 18:15:30 +1100:
 > 'Case numbers' is a meaningless metric, because it's impossible to 
know what each day's count means, it adds apples and oranges, and none 
of it tells anyone anything useful.

Editing some text from my email of Fri, 3 Apr 2020 11:49:04 +1100:
 > Countries have adopted very different approaches to recognising 
cases, and have changed their approach over time and space, sometimes 
frequently.
 >
 > Mostly, the sample of the population that is being tested at any 
given time is intentionally not random, but targeted.
 >
 > But the basis of the targeting (the sampling frame, and the manner in 
which the sample is selected from the sampling frame) is highly 
variable, and the execution of it is challenging and highly error-prone.
 >
 > One result is that within-country case-counts aren't comparable over 
even short periods, let alone the whole 4-8 weeks to date.
 >
 > A second result is that inter-country comparisons are completely 
meaningless, because the confounding variables dominate the data.


It suits the media to keep reporting numbers and dressing them up in 
graphics of various kinds.  And, to bolster the credibility of their 
graphics, they need to boost the mythology of modellers.  But the 
informational value of case statistics is at about the same level as the 
'what the celebrities did yesterday' pages.


Hopefully the policy-makers know all this, are ignoring the simpletons, 
are taking into account insights drawn from multiple partial models that 
deliver bits and pieces of insights into segments of the whole problem 
(Bernard's "multiple, interconnected models fed by real-life, current 
data"), and are making progressive and adaptable judgement calls based 
on what they have available to them at the time.


-- 
Roger Clarke                            mailto:Roger.Clarke at xamax.com.au
T: +61 2 6288 6916   http://www.xamax.com.au  http://www.rogerclarke.com

Xamax Consultancy Pty Ltd      78 Sidaway St, Chapman ACT 2611 AUSTRALIA 

Visiting Professor in the Faculty of Law            University of N.S.W.
Visiting Professor in Computer Science    Australian National University