Estimating Covid-19's Rt in Real-Time
github.com
github.com
That model uses the EpiEstim R package to estimate R based based on the number of case onsets and a serial interval distribution. EpiEstim implements the methods of Cori et al: https://academic.oup.com/aje/article/178/9/1505/89262 which also uses Bayesian methods.
It looks like the linked model is based on Bettencourt & Ribeiro: https://journals.plos.org/plosone/article?id=10.1371/journal...
The bigger problem is that R varies across the population; if you're testing primarily old age care homes you'll end up estimating what R is in that community rather than estimating R in the broader (mostly untested) community.
I was talking about the issue of different jurisdictions testing differently.
See this study from China where the use of onset of symptom data allowed them to estimate R much more accurately: https://twitter.com/XihongLin/status/1236075174069440512
From the official WHO Report, page 7, you can see how recording onset date effectively allows you to "see" a week into the future: https://www.who.int/docs/default-source/coronaviruse/who-chi...
$ head -1 kane-2020-04-19-2020-04-19-17:31:23-CDT.csv
Age Range,Age,Case Count,City,Death,Reported Date,Sex,State,Symptom Onset
$ awk -F, '!$9 { t++ } END { n = NR - 1; printf("%s %s %.1f%%\n", t, n, t * 100 / n) }' kane-2020-04-19-2020-04-19-17:31:23-CDT.csv
233 607 38.4%
https://www.kanehealth.com/For example, Massachusetts recently (around the time of the spike in its modeled Rt) made an aggressive push to assess the spread of COVID-19 in assisted living facilities.
Also the claims of "Rif A. Saurous" are dubious, the first of many:
"… it’s deeply implausible on its face that WA had R=.24 on 10 April and R=1.4 on 17 April."
That's the week Easter happened and the mobility data from Apple in fact does show people in Seattle traveling more that week:
Another recent gem:
When Will The Riots Begin? - Marginal REVOLUTIONmarginalrevolution.com › marginalrevolution › 2020/04 › when-will-th... 4 days ago - The economy was doing well, Trump had charisma but was erratic and ... Meanwhile the elites are laughing about binging Tiger King on Netflix.
The other is an email from a person not disclosing a real name making dubious claims of which even the first is not fact checked. Further the site makes policy suggestions routinely.
One of those is much more likely to be arguing in bad faith.
I'm saying that the arguments made in one fell to immediate scrutiny, wanting to affect policy likely shows why though.
In stats there's a smell test, arguments made without numbers.
Low-pass filtering helps smooth the curves, but adds lag, of course. Centered low-pass filtering is useful but can only be applied retrospectively. There's no magic bullet for dealing with noisy data.
So the would-be-modelers should really be able to handle at least that! Another very common pattern is that the cases / deaths / was test performed / whatever data which is to be reported gets "stuck" and then released in a bunch at some moment of time, sometimes only once in some week.
It's really surprising to me that such simple truths aren't recognized.
$ head -1 rt.csv
state,date,ML,Low_90,High_90,Low_50,High_50
$ awk -F, '$1 == "ID" { printf("%s: %.2f - %.2f\n", $2, $6, $7) }' rt.csv | tail -7
2020-04-13: 0.07 - 0.54
2020-04-14: 0.10 - 0.59
2020-04-15: 0.32 - 0.92
2020-04-16: 0.50 - 1.16
2020-04-17: 0.51 - 1.17
2020-04-18: 0.47 - 1.11
2020-04-19: 0.32 - 0.94For those who don't know, these are the two founders of Instagram.
If you're concerned with personal risk, the biggest thing that matters is the current spread. Of course if everyone is concerned with that, nobody will be concerned until it is too late to prevent a large portion of us from getting it.
However if you're concerned with the trajectory, all that you care about is whether R is less than or greater than 1. For both 2 and 10 it will hit a large part of the population, all at once. But 1.01 and 0.99 is the difference between "we all get it" and "only a small fraction of us get it".