6 ms·
Estimating Covid-19's Rt in Real-Time
- impostervt 7y agoI found this site very helpful: https://rt.live/ https://rt.live/
- jdminhbg 7y agoI was strongly convinced by criticisms of that model leveled here: https://marginalrevolution.com/marginalrevolution/2020/04/our-best-people-are-working-on-this-problem.html https://marginalrevolution.com/marginalrevolution/2020/04/ou...
- mzs 7y agoThat is not an impartial source: https://marginalrevolution.com/marginalrevolution/2020/04/re-starting-the-economy.html https://marginalrevolution.com/marginalrevolution/2020/04/re... 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: https://imgur.com/a/bCK7O86 https://imgur.com/a/bCK7O86 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.
- jdminhbg 7y agoI'm sorry, but MR is a mainstream, reliable source. The author has a column in Bloomberg, previously in the NYT, and is widely cited. The rt.live model is very bad, and you should take those criticisms seriously.
- SpicyLemonZest 7y agoWhat is an impartial source to you - someone who has no other opinions on related topics?
- mzs 7y agoOne is just data, plots, functions, methods, links to a peer reviewed paper, and cautions where dragons may lie in its results. It makes no policy suggestions. 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.
- SpicyLemonZest 7y agoI don't think it makes sense to have a heuristic of summarily rejecting any analysis from people who make policy suggestions.
- mzs 7y agoI don't either, and that's not what I'm saying. 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.
- mzs 7y agoalso links to this plus some informative blog posts
- Animats 7y agoThey're over-fitting noisy data. Huge changes in R over the course of a week, as for Idaho, has to be an illusion. One problem is that some states and some hospitals don't report daily data on weekends, and new cases and deaths are charged to Monday. 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.
- acqq 7y agoThe cases not being "perfectly spread" over the seven week days is observable also outside of the U.S. It's simply an artifact how most of the workers work. 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.
- mzs 7y agoIt does do smoothing, plus the ranges are very plausible for this Bayesian approach: $ 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 - 0.94
- lanewinfield 7y agoFYI, this github link and rt.live are both by Kevin Systrom.
- toomim 7y agoAnd rt.live was made in collaboration with Mike Krieger. For those who don't know, these are the two founders of Instagram.
- shalmanese 7y agoOne big limitation right now is that we're basing estimates on confirmed test positive date as opposed to first onset of symptoms which is a much cleaner signal. Unfortunately, first onset of symptom data doesn't appear to be collected anywhere in a systematic fashion. 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 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-china-joint-mission-on-covid-19-final-report.pdf https://www.who.int/docs/default-source/coronaviruse/who-chi...
- mzs 7y agoIndeed, for example I-NEDSS lacks an onset date for nearly two of five confirmed cases in Kane co: $ 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/ https://www.kanehealth.com/
- standardUser 7y agoI don't see how we can meaningfully estimate the reproductive rate with such variable testing rates and patterns. Some places barely test people with severe symptoms, other places are testing anyone they can that may have been exposed. Some states have tested less than 0.5% of the population, others have tested 3% plus. Testing criteria and rates and test availability are all changing from place-to-place in ways that I have not seen sufficiently tracked. So we can't even correct for those variations if we tried.
- cperciva 7y agoTesting rates and patterns don't make as much difference as you might think. They'll dramatically affect the absolute numbers, but to a first approximation the curves will have the same shape and thus estimates of R won't be thrown off. 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.
- rwcarlsen 7y agoThe curves won't have the same shape if test availability and circumstances under which people are tested change. And I think those things are definitely changing over time.
- cperciva 7y agoSure, if testing changes, it throws things off. (Unless you can figure out how to adjust for the changes in testing rate and methodology.) I was talking about the issue of different jurisdictions testing differently.
- mzs 7y agoThis is precisely where Bayesian stats shine: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0002185 https://journals.plos.org/plosone/article?id=10.1371/journal...
- korethr 7y agoA curious thing I notice in the final plotted graphs. Some of them show a continuous smooth decline, and when the trend dipped below 1, it continued to do so in a smooth fashion. Others, it dipped below 1, then bounced back up again, sometimes doing this a couple times. Assuming this is not a result from errors in the calculation or inaccuracies in the data, I wonder what the cause of this is. Was there some situational or behavioral change in those states where the trend bounced back up, vs. those ones where it continued to smoothly decline?
- twoodfin 7y agoMy (totally uneducated) guess is these spikes represent changes in testing patterns for which the model is not adequately compensating. 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.
- mindslight 7y agoDoes it strike anyone else as a bit weird to focus on R? It's obviously useful to know whether R is less than or greater than unity (subcritical or supercritical). But when R is greater than 1, then the growth rate over time seems more fundamental.
- dboreham 7y agoPerhaps the idea is to inform policy decisions, both government and personal. If I know the R0 is 10, I'm probably not going to leave my house. Otoh if it's 2 then with an N95 mask I'd think about grocery shopping.
- mindslight 7y agoExcept if that hypothetical R=10 is a slow acting virus that infects others over the course of a month, and the R=2 kills within a day...
- btilly 7y agoTo think this way is to misunderstand exponential growth. 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".
- baq 7y agowell, if it has very high R and hospitalizations/deaths are what they are, it's actually good news - herd immunity will come much quicker, possibly avoiding future waves, assuming immunity persists for a few years (which I hear is a reasonable assumption).
- AbrahamParangi 7y agoMight be worthwhile to estimate a per-state transfer function from deaths to cases to adjust for differential case reporting/testing. Deaths may be a clearer signal (though shifted and spread in time).
- Mvandenbergh 7y agoI rather prefer the approach here: https://epiforecasts.io/covid/posts/national/united-kingdom/ https://epiforecasts.io/covid/posts/national/united-kingdom/ which does not attempt to calculate r for the most recent days and shows days for which uncertainty is great in progressively lighter colours. Part of the excellent work done by the London School of Hygiene and Tropical Medicine's Centre for Mathematical Modeling of Infectious Diseases much of which can be found here: https://cmmid.github.io/topics/covid19/ https://cmmid.github.io/topics/covid19/. 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 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.pone.0002185 https://journals.plos.org/plosone/article?id=10.1371/journal...
- mzs 7y agoVery nice R, thanks! https://github.com/thibautjombart https://github.com/thibautjombart https://github.com/cmmid https://github.com/cmmid