1/ Yesterday I posed a question here on twitter after reading one of the daily updates by @bob_wachter (I love those updates!) The question was: why should the peak of the COVID19 epidemic come later when the curve is flatter. Should it? https://twitter.com/pleunipennings/status/1248835651732029440
2/ Specifically @bob_wachter wrote: "Having flattened curve, CA’s projected peak is later [than USA], April 15, w/ 1,616 total estimated deaths thru August."
– it sounded like he meant this to be a logical result: flatter curve --> later peak.
3/ I know why in epidemics that ultimately infect many, we expect earlier higher peaks (w high R) and later, flatter peaks (w lower R). I was one (of many) who tried to teach the world about flattening the curve. https://vimeo.com/396866214 
4/ But ... we are now in a different scenario. The stay-at-home /shelter-in-place orders are extreme and instead of a slow increase of the cases to reach a peak months from now, in many places we saw a decrease in the # cases about 10-14 days after the stay-at-home order.
5/ Here I show a clear example from my home country, The Netherlands. Schools closed on March 16th. The peak hospitalizations occurred 7-12 days later. Data from @rivm .
6/ We also see this in NYC. Peak hospitalizations 8-12 days after Stay-at-home order (on March 22).
7/ The fact that these curves are "bending down" (rather than flattening) is really great. The stay-at-home-orders and lock-downs and whatever they are called are working and saving lives.
8/ But that these curves are bending down is due to many staying at home and keeping a distance which leads to a sudden drop in R (how many people get infected by each case).
9/ The "normal" math of epidemics (flatter curve --> later peak) likely no longer holds under the current situation, because it depends on a slow decrease in R due to a build up of herd immunity.
10/ In the "classic" situation, the epidemic grows until R0*S < 1, where S is the fraction susceptible in the population. In a flatter curve, it can take longer to reach R0*S<1, which is why they peak later.
Now that we have reduced R drastically by staying at home as much as we can and not b/c R0*S<1, there is no longer reason to expect that lower curves peak later. Instead, curves would be expected to peak around 14d after a Stay-at-home-order (for hospitalizations).
12/ is this what we’re seeing?
13/ So if there is no reason to expect flatter curves to peak later, why does the @IHME_UW model predict this? The answer is not clear, and I don't know much about those models.
14/ But several colleagues reminded me that the @IHME_UW is a model that basically takes data, fits a curve and then projects into the future. This, in principle, is not wrong, but can lead to wrong predictions. @CT_Bergstrom @magnusnordborg @sarahcobey @DanSRosenbloom
15/ If all you have is the data (and no real understanding of the underlying process), then projecting from a fitted curve is not a bad idea.
16/ Say, you run a hospital and you had 40 new patients 3 days ago, 50 two days ago and 60 yesterday, it'd be smart to expect 70 today and 80 tomorrow and 140 a week from now. But if your mayor all but closes the city, then your predictions for next week may not be very good.
17/ It is actually really hard to predict the future of an epidemic when the behavior of almost everyone is changing dramatically, and it is unclear whether any of the models is doing a very good job at this. (That doesn't mean that we shouldn't try though).
Finally, I would like to suggest two reasons why a lower curve may actually really peak later. 1. people may be less strict about social distancing when there are few cases – meaning that R is not as much reduced.
19/ 2. Another is that when you have few cases, relatively many of these may be coming from travelers from elsewhere, which may be harder to stop than transmission within the city.
20/ is this why the curve in SF is not yet bending down?
Data here for hospitalizations @SF_DPH
You can follow @pleunipennings.
Tip: mention @twtextapp on a Twitter thread with the keyword “unroll” to get a link to it.

Latest Threads Unrolled: