Nate Cohn

Nate Cohn @Nate_Cohn

@delaneypage_ PA counts here — PolitiTweet.org

Nate Cohn @Nate_Cohn

@doug_rivers the 'raw average' here, btw, is the raw average of all other public polls, not the raw average of the microdata used, if that's part of the confusion here — PolitiTweet.org

Nate Cohn @Nate_Cohn

One of the odd things about the Trump effort to suppress absentee mail voting is that it's totally conceivable to me that higher turnout would help him in these Midwestern states, where the kind of lower turnout and less educated voters who sat out the midterms probably tilt-GOP — PolitiTweet.org

Nate Cohn @Nate_Cohn

@A_agadjanian @cwarshaw Sorry on all caps w there. In any case, the post LV screen yields a younger and whiter electorate IIRC — PolitiTweet.org

Nate Cohn @Nate_Cohn

@doug_rivers It’s basically just Ghitza/Gelman paper with a fixed effect more useful than last pres for a predictive setting. Hardly groundbreaking but not useless — PolitiTweet.org

Nate Cohn @Nate_Cohn

@A_agadjanian @cwarshaw yeah, my version above also using the CPS, and you’ll note some similar patterns (GA) due to some long standing biases in black turnout THERE — PolitiTweet.org

Nate Cohn @Nate_Cohn

@kwcollins @PollsAndVotes (and as @A_agadjanian observed, it does seem like the trend holds online, even if there's an intercept shift dating back years) — PolitiTweet.org

Nate Cohn @Nate_Cohn

@kwcollins @PollsAndVotes anyway i'd say this would be more in the category of 'death of telephone polling' than 'differential non-response bias' if it's wrong — PolitiTweet.org

Nate Cohn @Nate_Cohn

@kwcollins @PollsAndVotes iirc biden led among seniors in all but one cnn/ssrs, abc/post and quinnipiac poll going back to last summer — PolitiTweet.org

Nate Cohn @Nate_Cohn

@cwarshaw @A_agadjanian struggled more in '08 https://t.co/L834zUKYe0 — PolitiTweet.org

Nate Cohn @Nate_Cohn

@cwarshaw @A_agadjanian not the prettiest slide i ever made (filling in for mr. enten dropping out of something at the last minute, IIRC!) but here's a '16 effort that tried to synthesize a regression based on high-quality state polling and MRP w national RDD https://t.co/KBPkssEjZu — PolitiTweet.org

Nate Cohn @Nate_Cohn

@A_agadjanian and iirc you won't pick up almost all the changes you really care about. you're pretty dependent on the 'last presidential election' fixed effect in a predictive setting, and you don't have nearly enough n to figure out that IN will shift 20 pts to Obama but not OH — PolitiTweet.org

Nate Cohn @Nate_Cohn

@A_agadjanian i've done something like it going back to '04, though unpublished. regularization creates some persistent biases in mrp, at least using the covariates available from RDD surveys. state subsamples don't really get big enough unless you get dozens of national polls — PolitiTweet.org

Nate Cohn @Nate_Cohn

@b_schaffner depends on how bad your sample is--and i think some of the ivr/cheap online stuff is bad enough that even the exits help--but yes for most high-quality work — PolitiTweet.org

Nate Cohn @Nate_Cohn

@BrianStryker @kwcollins i think vast parts of the non-aapor, non-twitter state polling world are totally in the dark — PolitiTweet.org

Nate Cohn @Nate_Cohn

Notice, btw, that there's no big difference on party ID. These polls are weighted on party registration/primary vote history where available. And yet you can have a very meaningfully different result — PolitiTweet.org

Nate Cohn @Nate_Cohn

Similarly, our NYT/Siena polls in October would have been Biden+6 (rather than +2) without education or turnout weights (like most state polls) https://t.co/OSyhp4NzNc — PolitiTweet.org

Nate Cohn @Nate_Cohn

But take the national polls in '16. They would have been Clinton+8 without education weights https://t.co/FaV1BaodOt — PolitiTweet.org

Nate Cohn @Nate_Cohn

Now, that doesn't mean that every poll without education weights is off by 4 points from the 'truth.' It means that is' 4 points off of what it *would have been* if it had been weighted by education. Maybe a poll without cells or edu weights might post an 'accurate' result! — PolitiTweet.org

Nate Cohn @Nate_Cohn

It'll vary from state to state and pollster to pollster, but I'd guess it's around 3 or 4 points https://t.co/lVpe7u1cVV — PolitiTweet.org

Nate Cohn @Nate_Cohn

It'll vary from poll to poll and pollster to pollster. But I'd guess it averages out around 3 or 4 percentage points https://t.co/lVpe7u1cVV — PolitiTweet.org

Nate Cohn @Nate_Cohn

RT @ForecasterEnten: CNN poll standards as written by the smart @jennagiesta, who is our polling director, asks pollsters "If your survey h… — PolitiTweet.org

Nate Cohn @Nate_Cohn

@b_schaffner yeah it is a sad world :( — PolitiTweet.org

Nate Cohn @Nate_Cohn

@Onlypizza1 @jonathanchait @ForecasterEnten yup — PolitiTweet.org

Nate Cohn @Nate_Cohn

@RachelBitecofer @ForecasterEnten surely you must realize, despite the snark, that errors due to education weighting and due to a bad turnout model (say, weighting to a '14-like electorate in '18) are entirely independent — PolitiTweet.org

Nate Cohn @Nate_Cohn

@kwcollins and to clarify, by lack of knowledge, i mean 'they don't know anything about this whole debate. they don't know about education's role in '16. they don't realize they're doing something controversial. they're just doing what they've been doing, with no thought' — PolitiTweet.org

Nate Cohn @Nate_Cohn

@swiley1983 @szorowitz @jonathanchait @ForecasterEnten https://t.co/8jxraB3hJb — PolitiTweet.org

Nate Cohn @Nate_Cohn

@MasabuAida @hezaproject @kwcollins and most of these polls don't contact unregistered voters (CVAP probably close to RVs, but in fairness to them it's not the same thing so I can understand the reluctance) — PolitiTweet.org

Nate Cohn @Nate_Cohn

In high-quality live interview telephone surveys, weighting to the exit poll is often worse than not weighting at all. That may not be the case here — PolitiTweet.org

Nate Cohn @Nate_Cohn

@kwcollins lack of knowledge, bad/uncertain targets (exits being a big problem here), obstinance + belief that it didn't ruin their prior results — PolitiTweet.org