Several people have expressed their surprise, sometimes more than that, that I keep harking on this, so here is a short thread to explain why I don't let it go and why I won't until people admit this model was completely wrong. 1/n
First, people like to say that science is self-correcting, but this doesn't mean it happens magically. Someone still has to do the correcting and, right now, epidemiologists don't seem very interested in doing it. 2/n
It's a real problem that, purely out of tribalism (the people who criticize the ICL model tend to be on the "wrong" side of the political divide), many of them continue to defend the model by using the demonstrably flawed argument I debunked above. 3/n
But another reason, which is perhaps even more important, is that since people won't even admit the model was wrong, there is no discussion of *why* it was wrong. 4/n
People just say that it's because it didn't model voluntary social distancing and move on saying "nothing to see here". But again this is demonstrably *not* the reason, since it's still wildly off the mark even when you include voluntary social distancing in the scenario. 5/n
Based on what I have seen when I ran a modified version of the model for Sweden, the proximate reason seems to be that, in the model, the virus spreads way faster than it does in reality. 6/n
For instance, even when I conservatively model voluntary social distancing based on Google Mobility data, it predicts that ~50% of the population has already been infected by now. 7/n
Now, I don't know exactly what proportion of the population has already been infected in Sweden, but I have no doubt that it's nowhere near 50%. So why does the model get the dynamic of the epidemic so wrong? 8/n
This is a really interesting and important question, but since people are too busy pretending the model wasn't really wrong, nobody is really trying to find out... 9/n
I don't really know why the model got the dynamic of the epidemic so wrong, but I have a few theories. One of them is that it's because, although the model assumes they are, people are not in fact equally susceptible to infection. 10/n
A team of epidemiologists has recently shown that variation in susceptibility could have a huge impact in a SEIR model. I suspect this could also be the case in an agent-based model such as ICL's. 11/n medrxiv.org/content/10.110…
This wouldn't be very difficult to test by making a few chances in the code and maybe I'll do it when I have some time, but right now I don't and I don't know when or even if I will. 12/n
I'm also not in the best position to do this, because my laptop was clearly not made to run such very computationally intensive simulations. It took me 65 hours to run the simulation I talked about above 🙃 13/n
This is not the only possibility, I can think of others and I'm sure there are many I haven't even thought about. I think understanding why ICL's agent-based model got the dynamic of the epidemic wrong could be very illuminating. 14/n
Unfortunately, for this to happen, people first have to admit that it got the dynamic completely wrong and not pretend that it's just because they didn't run a scenario that included voluntary social distancing, which is total bullshit. 15/n
So until people acknowledge that and stop repeating this bullshit, I won't stop harking on that. And I think people who say it's not important and that we should just let it go are extremely wrong. 16/n
What I said doesn't exhaust the reasons why I think it's important not to pretend there is nothing to see here, but I need to go back to work, so I'm going to stop. 17/n
It's frankly amazing that I even have to defend myself for not letting people get away with shoddy science. The truth is that, had this model been used to justify policies you disagree with, I'm pretty sure you wouldn't be singing the "nothing to see" tune right now 🤷♂️ 18/18
ADDENDUM: I agree with the gist of this thread, i. e. there is no obvious way to translate Google Mobility data into interventions for the model, but I actually thought about that and that's also not why the model got the dynamic of the epidemic wrong.
Basically, since I knew that people would make that point (which again is totally fair), the first thing I did was not run the above simulation but a scenario with a hard lockdown, all schools and non-essential business closed, etc.
But the model still predicts the virus spreads much faster than it actually does *even in that scenario*, although obviously not nearly as much. 
Now, I agree that it's not obvious how to translate Google Mobility data into interventions for the model and that reasonable people can disagree about the right way to do that.
But I *don't* think reasonable people can disagree that there was a lot more social distancing in the hard lockdown scenario I ran than what actually happened through voluntary social distancing.
So while it's clear that, depending on how you translate Google Mobility data into interventions for the model, it may not look as bad, it will still looks pretty bad on *any* reasonable way of doing it.
Like I said when I first tweeted about the simulations I ran, I plan to write something where I explain in more details what I did. One thing I plan to discuss is precisely the "translation problem", but also why it can't save the model.
If I have time, I will try to run one or two other scenarios, to cover a wider range of the different ways in which one could translate Google Mobility data into interventions for the model. But this will only strengthen my point.
By the way, the full lockdown scenario "only" took 37.5 hours to run on my laptop, because less people get infected and the loop on infected people is apparently the most computationally intensive part of the simulation, so you can try it at home 😄
There are other things I didn't explain. For instance, the "actual" curve actually accounts for reporting delays, because if you use the raw data it will look as though deaths are dropping faster than they are. Again I will explain all that when I write about it.
