Re: Sage & onion
I'm a bit worried about the trend towards criticism of epidemologists and public health modellers generally based on micro-analysis of small samples of particular implementations of models.
Well, if there are enough models, then there'd be scope for WHO to.. model ensembles on the WMO's work. Which is another interesting modelling challenge. On the one hand, we're told the 'science is settled', on the other, 100+ models producing widely different results. But having slowly standardising an approach, it makes reanalysis simpler. But both are predictive models, which can be 'fact checked' against reality.
People are playing defence lawyer when they should be looking at the basics (one person infects between 2.4 and 3.6 other people on average when there is no immunity in a (sub)population, there is a 6 to 14 day period when someone is infectious but not showing symptoms, simple differential equation)
But that isn't what the Ferguson model claimed. The claim was 500k dead, and that's looking unlikely. But then there was also a previous prediction of 200m dead globally from bird, swine & BSE, where the reality was <1,000*. Or one of the odd situations where mortality was exceeded by '57 & '68 flu outbreaks. Or some other bad years, where the BBC & rest of the MSM didn't run screaming headlines and running death counts.
So R0 is just one parameter that is/was uncertain.. Which model predictions translated into lockdowns to minimise impact if R0 is high... But if mortality rates are low, then lockdowns may have been the wrong approach, and 'herd immunity' more effective. In an actuarial sense, once the panic is over, it'll be possible to look the cost per life. The kind of grim calculations insurance types (and politicians) are expected to do.
This is applied (messy, political, entangled with desperate economic imperatives) science not systems level software engineering. The basis of science is that independent work by different teams using different methodologies tends (with eddy currents) to converge on something that we can regard as close to truth or reality in some sense. This is not an argument about engineering change orders or which library to use or what code style might be best.
It probably should be, given many scientists aren't software engineers. That's just a necessary evil.. I mean skill they'd need some familiarity if they want to analyse data.. Which they can frequently get wrong. Politics comes in when politicians decide to act on bad data, so where this is critical, perhaps more formal methodolgy should be used. I mean I was forced to learn Z at Uni, so some scientists should also be expected to suffer!
Again there's also a risk from consensus science. That's been pretty common in climate science, especially if anyone dares to try and publish something that challenges the consensus. A crucial variable there is CO2 sensitivity, and if you assume that's high, models predict high warming rates. But frequently fail reanalysis, ie comparison against historical data. If CO2 sensitivity is low, so will warming rates... But that's a rather controversial view because that 'science' has become highly politicised.
Same challenge exists with epidemeological models. Assume high R0 and mortality, it'll predict a lot of deaths. Reality may disagree, but that's where scientists need to try and educate politicians about confidence levels.. And ideally get the ONS to explain statistical confidence tricks that can get used to create false confidence. Especially given the money at stake for treatments and vaccines, like one of the previous trials where the sample size was very small, and the outcome had to be redefined to show any significance at all. Then there's stuff like 'relative risk'. Take 2,000 patients, give 1,000 a placebo, another 1,000 balm from Gilead. Placebo group has 2 deaths, balm group 1. Claim treatment offers a 50% improvement in outcomes rather than a 0.1% difference. But those kinds of stats sold an awful lot of statins..
* Swine flu in India may have killed many more, depending on how that's defined.
Also found another interesting article-