(yay!š©)
1/n
I'll try, for once, to post a few tweets that are as politically & ethically neutral as possible (but I still think fascists suck!) and only deal with *NUMBERS* (thanks @andhans_jail for the gif).
For some reason, most of my Twitter timeline seems now full of a weird version of "Price Technical Analysis", taking random data-points, projecting lines (often curve, sometimes ECCCCHXPONENTIAAAAAL!!!!), extrapolating magical functions, predicting trends, etc.
I will try not to directly discuss this tendency, here (even if I could, since many people, including me, may consider this kind of forecasting as too unreliable to be used as a foundation for important personal strategies & choice on it, ...
... let alone using it as a foundation for reckless central-planning experiments with play with the life of billions of different people interacting in millions of unknown contexts, to be enforced via violent enforcement by men with guns, badges, funny hats and sunglasses).
You guys seem very excited about trends in the "number of cases" you read about on some nice spreadsheet online. But that is just (and I'm so exhausted from repeating it for a month at this point) THE NUMBER OF TESTS EXECUTED TESTS WICH TURNED OUT POSITIVE (plz read twice)!!!
Before you full-time exponential-experts get all triggered because I'm ruining your noble efforts to save the life of millions by teaching them the secrets of exponentials and the art of staying home (& occasionally supporting Orwellian dystopias): that's not my point here!
Indeed, my point is that number is seriously *underestimated* as a proxy for real cases! Even worse: it's *arbitrarily* underestimated, in a context where the incentives to use it as a political tool (to show that some policy is "urgent", or not, or that it "worked", or not)!
As a matter of fact, I think that number was (ab)used a lot kinda everywhere (with the notable exception of South Korea) to confirm the bias of a very localized & containable contagion, which could be tracked & managed at an individual level. This was achieved (let's ...
...ignore right now how often intentionally: not my main point here) by initially testing (almost) only people with clear & obvious links with preexistent epidemiological assumptions/conjectures, including *debunked* ones, as in the case of the Italian "patient 0" (sic).
Then the poor number was (ab)used again by some governments to "prove" that the territory they control was magically immune, & by some others to "prove" that any kind of political intervention was "urgent" (or, maybe, "useless", depending on politically preferred narratives).
Not only the underestimation comes from testing *not* statistically representative/neutral samples (initially the bias was about simple/clear epidemiological assumptions, now it's about the severity of symptoms): the swabs have themselves a false-negative rate of about 2/3!
Of course, this severe underestimation can, in turn, cause a severe *overestimation* in the fatality rate of a disease. Within a fraction, if your denominator gets artificially reduced, your result will be artificially increased. I don't think it's a coincidence that ...
... in al the cases where the testing was statistically meaningful (South Korea again, but also quarantined cruises, and the little town of Vo in Italy, or some micro-nations where it was feasible to test the entire population) the fatality rate was (very much) lower.
Actually, there's another, more indirect reason, for the overestimation of the fatality rate when the number of infects gets seriously underestimated: since the number of "covid19 deaths" counts all the positive-tested people who happen to die, regardless of the actual ...
Do you really want a "death toll" which is mostly bias-free, & still informative enough in case of "apocalyptic" scenarios? Here, use this (>65 yo total deaths, by country, in Europe).
euromomo.eu/outputs/zscoreā¦
Look for Italy, for example. Almost no selection bias possible.
Of course this will be more reliable next week (data for the 11th week is still coming in: they are adjusting for delays). But it's already incomparably more reliable than completely arbitrary/gameable test-based metrics! In a couple weeks the trend should be quite obvious.
Say what? You can't wait weeks to take important decisions for your health & that of your (especially older) loved ones? Especially in "EXPONENTIAL GROOOWTH!!" scenarios where it's crucial to act before you have the actual numbers? That's fine: predict, forecast, anticipate!
That's what I'd suggest you do *in general* (it's what I do, trying my best to take care of my family: we should all be safe rather than sorry). But I'd actually *urge* you to do it if you use your predictions as a base for advocating centralized, violent social engineering!
Also, lags! The vidence seems to suggest average time from contagion to 1st symptoms is 5 days, from those to death 20 days. The very first lockdown in Europe (in Italy, yay!) literally happened 11 days ago! Sure, there's some variance, some faster outliers. Sure, for ...
... older people the latter is more like 13 days. Still: 18>11 & averages matter. Don't do "morgue TA" ignoring lags. Or do it, if you like it, but don't use it to assess how much some of your favorite fascist policy "iS WoRkiNg BeTTeR!!!", please.
Ok.
FIN.
There are a few typos. But you guys don't pay me for tweeting. So, live with it. š