Nate Silver

Nate Silver @NateSilver538

Assuming there's a slight lag between when these tests were conducted and when the results were published, and taking everything else at face value, we'd be looking at a ~13-15x undercount in NYC and maybe more like ~8-10x in NY outside NYC. — PolitiTweet.org

Nate Silver @NateSilver538

Many caveats here and we have to learn more about how the tests were conducted, which could bias the numbers either upward or downward (or some mix of both). But, interesting data. — PolitiTweet.org

Nate Silver @NateSilver538

But if I go to a movie theater with 100 others and it gets spread there? We moviegoers don't know one another. We didn't talk before the movie. We won't talk afterward. It would likely just look like generic "community spread". — PolitiTweet.org

Nate Silver @NateSilver538

It seems likely there is bias in which superspreader events survive into the fossil record. If I and 100 other people attend a wedding and a number of us get sick, we'll talk to one another afterward (wedding guests know one another!) so we'll know it was spread at the wedding. — PolitiTweet.org

Nate Silver @NateSilver538

It is *likely* that the situation in California is improving, but the state's data is such a mess that I'm not sure we'd really know if it wasn't. — PolitiTweet.org

Nate Silver @NateSilver538

California has some of the worst data in the country. Substantial lags in reporting, very hard to get a handle on its testing situation. — PolitiTweet.org

Nate Silver @NateSilver538

Jordan, Stockton, Miller, Pippen, Rodman, Robinson, starting lineup depends on matchups — PolitiTweet.org

Nate Silver @NateSilver538

In states that *are* diligent about reporting negatives, like New York, you *can* literally say "here's what share of tests were positive yesterday". That doesn't work in states like California where negatives and positives are reported with different lags. — PolitiTweet.org

Nate Silver @NateSilver538

Importantly, the negatives do not necessarily line up with the positives. If I take a test Tuesday, and it's positive, it might show up in the data Wednesday in California. If it's negative then, who knows—it might show up next week. — PolitiTweet.org

Nate Silver @NateSilver538

Instead, for states like these, it means "we've had 150K more negative tests results since we last could be bothered to give you a good estimate of negatives". Odds are that many those negatives actually occurred days ago or even weeks ago. — PolitiTweet.org

Nate Silver @NateSilver538

MANY states, California being the perhaps the worst offender, report positive tests diligently but negative tests erratically. So when you see a state report 150K "new" negative tests, it doesn't mean the tests were all conducted yesterday. — PolitiTweet.org

Nate Silver @NateSilver538

On the cases and tests: California reported a huge backlog of negative tests today, which is why there was the sharp one-day increase. It probably does not represent a new normal of higher testing volumes, unfortunately. — PolitiTweet.org

Nate Silver @NateSilver538

First, on the deaths: good that the numbers are down not only from the awful ones yesterday, but also relative to a week ago. Leaves hope that some of the #'s yesterday were states catching up on deaths that had been missed before. Still: over 2K per day so not much to celebrate. — PolitiTweet.org

Nate Silver @NateSilver538

US numbers via @COVID19Tracking. Weird but better day. Newly-reported deaths Today: 2,108 Yesterday: 2,674 One week ago (4/15): 2,492 Newly-reported cases T: 27K Y: 27K 4/15: 30K Newly-reported tests T: 311K Y: 152K 4/15: 161K Share of positive tests T: 9% Y: 18% 4/15: 19% — PolitiTweet.org

Nate Silver @NateSilver538

On Super Tuesday (3/3), maybe the last normal-ish day in American politics before COVID became the only story, Trump's approval rating was 43.3 and his disapproval was 52.7. Now? 43.6% approve and 52.4% disapprove. So he's lost his small bounce. https://t.co/Vfmzd6B2ps — PolitiTweet.org

Nate Silver @NateSilver538

But if you have some prior that the underlying data should be reasonably smooth and continuous, and the main challenge is that the estimates you get of the underlying function are noisy/stochastic, they can be useful. — PolitiTweet.org

Nate Silver @NateSilver538

LOESS methods are often misused because they do not handle *true* discontinuities in the data very well, i.e. some huge event happens and there's a huge, essentially instantaneous shift in the polls. You *don't* necessarily want to smooth that over. — PolitiTweet.org

Nate Silver @NateSilver538

Geeky note: LOESS-type smoothers are often overused and sometimes even abused. But they're a good use for COVID-19 data where there are fairly consistent trends on a ~weekly basis but also erratic reporting from day to day. — PolitiTweet.org

Nate Silver @NateSilver538

Hope this was a really useful conversation for both fans and critics of the @IHME_UW model. — PolitiTweet.org

Nate Silver @NateSilver538

@kmedved Yep, that makes sense. — PolitiTweet.org

Nate Silver @NateSilver538

So a 6-to-1 ratio of infections to detected cases in Geneva and a 30-to-1 ratio in California might actually mean basically the same thing, given the high rate of testing in Geneva and the low rate in CA. — PolitiTweet.org

Nate Silver @NateSilver538

To take another relevant example, NY state has done about 4x more tests per capita than California. (California's testing situation is poor BTW and that should probably figure into some of the narratives that are praising leaders over there for their response.) — PolitiTweet.org

Nate Silver @NateSilver538

In Geneva canton, for instance, there have been about ~45 tests per 1,000 people. That compares to ~25 tests per 1,000 people elsewhere in Switzerland, or ~13 tests per 1,000 people in the US. https://t.co/IXgpLYmDZm — PolitiTweet.org

Nate Silver @NateSilver538

In looking at these serology studies, which on the surface may seem to show much different results from one another, it's important to account for how much testing a given region is doing. — PolitiTweet.org

Nate Silver @NateSilver538

@kmedved @DanRosenheck It may have evolved a bit but the methodologies that produce super high R's are often dubious IMO (i.e. not accounting for increase in detections) and anything much greater than 3 (perhaps outside of relatively narrow settings) would show up in the deaths & hospitalizations data. — PolitiTweet.org

Nate Silver @NateSilver538

@DanRosenheck It's also extremely dubious if he's claiming that R is 3-5, which is definitely not the consensus view. — PolitiTweet.org

Nate Silver @NateSilver538

Deaths are not necessarily a perfect way to estimate infections (the IFR may vary from place to place for many reasons) but they're at least decently good. So it could *easily* be the case that say 30% of the population had COVID-19 in NYC but only 1.2% did in Santa Clara. — PolitiTweet.org

Nate Silver @NateSilver538

As a related thought: I think people underestimate how much more prevalent COVID-19 is in some areas than others. NYC has around *30 times* more deaths per capita than Santa Clara Co., Calif. — PolitiTweet.org

Nate Silver @NateSilver538

On the flip side, this is much less of a concern in areas with a greater underlying incidence in the population. If an area has been 15% infected, then a potential false positive rate of 1% or 2% or 3% isn't going to matter as much. So I'd look at those studies instead. — PolitiTweet.org

Nate Silver @NateSilver538

There are a lot of well-intended and well-written critiques of the Santa Clara Co. serology study but at some point it's not that complicated. A test that *could* have a false positive rate of up to ~2-3% isn't saying very much if it detects 2-3% positives in some population. — PolitiTweet.org