Nate Silver

Nate Silver @NateSilver538

RT @ABCPolitics: DACA is safe—for now. @FiveThirtyEight on what could happen next: https://t.co/TulcTh9LOQ — PolitiTweet.org

Nate Silver @NateSilver538

@kwcollins We use a prior based on the historical RV vs LV gap. Generally LV polls help Republicans but hurt incumbents, so a bit of a mixed bag for Trump this year. Then the prior is adjusted as we collect polls that provide both RV and LV results. — PolitiTweet.org

Nate Silver @NateSilver538

All right, that's it! If you want even more detail on the methodology, you can find it here. Also taping a podcast on this shortly, which we'll have out soon. https://t.co/FxbzezloWI — PolitiTweet.org

Nate Silver @NateSilver538

Our polling averages do *not* correct for that kind of thing. They're entirely based on the polls, albeit using the polls in some "fancy" ways. Our forecast model, conversely (not released yet) will blend the polls with other priors/methods to smooth things out a bit. — PolitiTweet.org

Nate Silver @NateSilver538

Right now, for instance, Biden is up 10 points in MI but 5 points in PA. I'm not sure I buy that they're really that different, or even that MI is necessarily the better state for Biden. But MI polling is often weird, and there hasn't been much high-quality PA polling lately. https://t.co/Waxnyug2r9 — PolitiTweet.org

Nate Silver @NateSilver538

The timeline adjustment helps a *lot* in giving you an accurate picture of the race at any given time, especially pre-Labor Day when polling is sporadic. Note, however, that there are still going to be states where there isn't much polling or the polls are of mixed quality. — PolitiTweet.org

Nate Silver @NateSilver538

This means our averages can move *even there hasn't been new polling in a state*. Here's Nevada, for instance, where there's little recent polling but the algorithm is making inferences based on changes in national polls, Arizona and Colorado polls, etc. https://t.co/nLrpwNdA0m https://t.co/5bscmBRWKN — PolitiTweet.org

Nate Silver @NateSilver538

Finally, we calculate a "timeline adjustment". Basically, this adjusts old polls based on trends in national polls and polls of similar states. (so, e.g. Wisconsin polls will have an influence on the timeline adjustment in Michigan polls, but Hawaii polls won't). — PolitiTweet.org

Nate Silver @NateSilver538

So house effects are now *mostly* calculated *within* states. I realize that this is a geeky technical point, but it makes a pretty big difference as far as accuracy goes. It gives local pollsters a proper amount of influence in the average. Local pollsters are often pretty good. — PolitiTweet.org

Nate Silver @NateSilver538

We've found that it can be a mistake to apply house effects too aggressively **across states**. Basically what happens is that only prolific pollsters that survey across lots of states wind up having much of a say in your average and the influence of local pollsters is reduced. https://t.co/Ht3h3gw9Fz — PolitiTweet.org

Nate Silver @NateSilver538

Next, there's a house effects adjustment. If polls consistently show good results for one candidate (i.e. a pollster always shows Biden doing 5 points better than other polls of the same state), then part of that house effect is corrected for. However, one needs to be careful... — PolitiTweet.org

Nate Silver @NateSilver538

Polls are adjusted in three ways. First, there's a likely voter adjustment. That is, polls of registered voters and all adults are translated to what we'd expect they'd say among likely voters. This helps Trump slightly, although there's likely not a huge RV vs. LV gap this year. — PolitiTweet.org

Nate Silver @NateSilver538

The trendline method is usually more aggressive, at least how we calculate it. This yields better projections at the end of the race, but it can be too "bouncy" earlier on. So we mostly use the averaging method now, but will mix in more of the trendline in October/November. — PolitiTweet.org

Nate Silver @NateSilver538

As for how we calculate the average, there are basically two ways to do it. You can either take a moving average (like RCP) or fit a trendline (like HuffPo Pollster). Our method is basically a blend of the two, which is a bit more accurate than either method on its own. — PolitiTweet.org

Nate Silver @NateSilver538

Polls are weighed based on their sample size and their pollster ratings. We really do try to be as inclusive as possible, rather than having to make a lot of judgment calls about which polls we like. But there are an increasing number of edge cases so we do have some rules. https://t.co/N2myuKuxz9 — PolitiTweet.org

Nate Silver @NateSilver538

We also have pages for individual states, once they've had enough polling. Here's Florida, for instance. Wisconsin, Michigan and Florida are three states where there's been especailly notable movement to Biden, even above and beyond the national trends. https://t.co/nWUo9X5F4y https://t.co/MDGVSqWpV3 — PolitiTweet.org

Nate Silver @NateSilver538

OK, a bit more on our polling averages. Biden leads by about 9 points nationally. This reflects recent movement. His lead was closer to 6 points at the end of May. https://t.co/cy51vc5isJ https://t.co/TTHpjJngaz — PolitiTweet.org

Nate Silver @NateSilver538

I see there's breaking SCOTUS news, so I'll save a longer thread for a bit later. But here's a story detailing what we're doing here. This is not our forecast model (so no probabilities yet, those will come later) but there's a lot of work in these. https://t.co/FxbzezloWI — PolitiTweet.org

Nate Silver @NateSilver538

Hey folks, we've got our general election polling averages up! These will update continuously through November. Joe Biden currently leads Trump by 9.2 points nationally. https://t.co/cy51vc5isJ — PolitiTweet.org

Nate Silver @NateSilver538

Probably also some effects due to weather, as mentioned before. If you control for the total number ever infected and google searches for "air conditioning" (as a proxy for people being indoors vs. outdoors) Trump margin of victory has no effect on current R. https://t.co/YTyvOZX56B — PolitiTweet.org

Nate Silver @NateSilver538

This graph compares the total number ever infected via @youyanggu versus current estimates of R via https://t.co/LwOqS5Fl1c. There is a strong (correlation = -0.69) inverse relationship. The *less* COVID a state had before, the *faster* it's spreading now. https://t.co/Hm4gKAPUyW — PolitiTweet.org

Nate Silver @NateSilver538

This dynamic is interesting but complicated. While it's plausible there's a relationship between political partnership and COVID spread, e.g. if Trump voters are doing less distancing, right now by far the best predictor of a state's R is how many people had COVID before. — PolitiTweet.org

Nate Silver @NateSilver538

In NYC, the death rate among people aged 75+ has been *40 times higher* as a share of known cases than among people aged 18-44. Similar numbers in Wuhan. So, even subtle shifts in the age distribution of cases can matter a lot. https://t.co/lWxCdAhq2J — PolitiTweet.org

Nate Silver @NateSilver538

There is increasing evidence that new cases tend to be younger than the ones in March/April, possibly essential workers or people returning to work or social activity after lockdowns. That helps to explain the seeming incongruity between deaths and cases. https://t.co/CX8mqopC3p — PolitiTweet.org

Nate Silver @NateSilver538

Typical for recent days. Deaths down week-over-week; the 7-day average has fallen to 659. But cases are *up* slightly. The 7-day average case count has risen to 22,582 from a low of 19,957 on 6/9. The positive test rate is roughly flat, meanwhile. — PolitiTweet.org

Nate Silver @NateSilver538

US daily numbers via @COVID19Tracking: Newly reported deaths Today: 782 Yesterday: 713 One week ago (6/10): 878 Newly reported cases T: 24K Y: 23K 6/10: 21K Newly reported tests T: 489K Y: 465K 6/10: 423K Positive test rate T: 4.9% Y: 5.1% 6/10: 4.9% — PolitiTweet.org

Nate Silver @NateSilver538

@joshtpm Sorta reflects Northeast media bias (only half-joking). — PolitiTweet.org

Nate Silver @NateSilver538

Would like to read more on if there's a relationship between where COVID is spreading and people being indoors more because it's too hot out. Here are searches for "air conditioning", for instance. Note that AZ (very high COVID growth now) tops the list. https://t.co/EJFrdimrJX https://t.co/AVImYydjXJ — PolitiTweet.org

Nate Silver @NateSilver538

The average election year features 10 full-width headlines through Election Day. There have already been *33* this year, and we're not done with June yet. — PolitiTweet.org

Nate Silver @NateSilver538

So, to get a sense for how much *news* there is in the a given election year, we looked at how many full-width headlines there are in the NYT from Jan. 1 through Election Day in election years going back to 1968. 2020 is, uh, pretty special... https://t.co/9HzXC0jxXK — PolitiTweet.org