So, the errors in the Imperial College covid-19 model. It sounds complicated, actually it’s incredibly simple.
Their model (which isn’t actually that complicated, uses standard libraries to do the heavy lifting, quite properly: [github link]) is fetching the data for cases and deaths automatically from the ECDC dataset.
The issue is, though, that for Sweden and the UK (at least), the death figures will be wrong since the ECDC figures aren’t the actual number of people who die that particular day: they are the update to the cumulative totals which are announced each day.
So for example, one day the UK might get a big batch of results back from the labs, which they then allocate back to the days those people actually died, and they then publish each day’s actual deaths on the NHS website with the updated number of people who actually died on each particular day. Sweden works the same.
But the problem is that Imperial is taking the daily updates to the cumulative number as each day’s death. This means firstly that the trajectories are wonky, and secondly also that they are working off the wrong “base” numbers – in the UK for example we had a day recently where 980 odd results came back, which obviously will have looked like a bad “death day” and which the Imperial number will fit its curve too and use as a springboard for the future, but really only 730-odd people *actually* died that day. Today we announced about 750, but only 113 actually died yesterday, the rest are deaths going back weeks – and that 113 will get pushed up over the next few days (particularly next 2 days) as post-mortems and late filings and etc. come in for people who died yesterday. Recently a lot of results have been coming back, but the people actually died quite a while ago.
So they’ve built a very slick chained-Markov Monte Carlo model that produces beautiful output but is based on totally the wrong numbers.
It’s actually not that difficult to do a better pro forma: for my version I take each day’s actual death figures (updated daily from the NHS), inflate say the last 5 days numbers by the average that those numbers tend to get pushed up (which is pretty consistent), and you have a much more reliable dataset. Not perfect, sure, but better. At the pretty simple end of data wrangling.
Averaging this across all the countries in a big dataset would be fine, but it would be nice to try and get the data right for countries that are making policy decisions based on this model (I’ve been in lockdown for 3 weeks now largely off the back of it) but I’m whistling into the wind.
Haimes illustrates the difference between modeling the reporting process, which we’ve been doing, and modeling actual deaths (and cases etc.). It just isn’t the same thing.
A reporting model is useful for political projections, by which I mean as fodder for guessing what our elites and leaders will do. And what a panicked populace does.
An actual-numbers model is useful for politicians and elites to make plans. By which I mean, in what they will do to us.
Obviously—or I hope obviously—a reporting model doesn’t have to get the actual numbers right. It doesn’t even have to know anything about the actual numbers. It does have to (or should) get the reported numbers right.
But an actual-numbers model must get the actual numbers right, or else the plans our leaders make will be based on error. And therefore it’s likely the plans will be suboptimal.
Our strategy of PANIC! PANIC! PANIC! can’t be considered by optimal by anybody who is not a politician or elite seeking greater control.
Maybe you’re skeptical about the political control bit, preferring not to be as black of heart as your host. If so, take a look-see at this headline: US may have to keep some social distancing measures until 2022: study.
Subtitle: “Harvard researchers used computer models to simulate how the COVID-19 pandemic could play out.”
There are still a few people amazed about models run on “computers”, as if the act of running models on a computer guarantees success. There are more who assign the same scrupulosity to Harvard products. (Yes, really.)
But this model is so asinine it doesn’t need refuting. Do these foolish academics really believe we have never had pandemics before and got through without worldwide government control? Don’t answer.
That is, answer this instead: do you think some power hungry elite will use Hahvahd’s model as an excuse to tighten her grip?
To support this site and its wholly independent host using credit card or PayPal (in any amount) click here