Nate Cohn

Nate Cohn @Nate_Cohn

I went back to the Times/Siena surveys from June. What do you think was the difference between the response rate among registered Democrats and Republicans? — PolitiTweet.org

Nate Cohn @Nate_Cohn

I went back to the Times/Siena polls from July. What do you think was the difference in response rate between registered Republicans and Democrats across the seven surveys? — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris @NateSilver538 'all else equal' is the hang up. the material question is whether MRP competes with topline-based averages/regression in the actual pre-election information environment. if we get the microdata for all the polls being aggregated, MRP could compete. that's not the world we have — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris @NateSilver538 seems totally unimportant? who cares whether there's any material advantage to average/regression of 10000 state polls from 1000 pollsters v. MRP with 10000 state polls from 1000 pollsters (there isn't) — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris @NateSilver538 but the point of nate's exercise is what to do when you don't have much/any state polling at all, let alone the microdata! i don't see how it's remotely equivalent — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris @NateSilver538 if by 'equivalent data,' you mean that you have microdata from a dozen national and state polls--that you don't get IRL--then that would certainly cancel out issues with regularizaiton, as I believe I said in the context of GA — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris @NateSilver538 not nationwide, that's for sure. and there's an example of why maybe you've got to be careful with using single polls. fox hasn't shown biden up by more than 10 points among white college graduates in their last 5 polls. good luck using that for this purpose — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris @NateSilver538 and there's an example of why maybe you've got to be careful with using single polls. they haven't shown. biden up by more than 10 points among white college graduates in their last 5 polls. good luck using that for this purpose — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris @NateSilver538 i don't have even half that much microdata! — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris @NateSilver538 i don't tbh, and this scenario supposes more microdata than anyone's actually ever had in a pre-election setting — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris @NateSilver538 i actually don't tbh — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris @NateSilver538 if by state-level microdata you mean 'the microdata from a N>400 or whatever state poll', then i would presume that your estimate in that state would be comparable to a single state poll — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris @NateSilver538 sorry; deleted for length and then it didn't make sense: 'certain it would not compete with state polling averages' — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris @NateSilver538 you can be fairly competitive with the state polls if you go beyond this, and either have far stronger covariates or a way incorporate state polling data (in a manner like other nate proposes) or pass along the state-by-state biases in MRP from prior elections — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris @NateSilver538 if by 'mrp' you mean a model with regularization, deeply interacted standard census demographics, and some standard political/demographic continuous predictors, then i am certain it would not--even with tens of thousands of RDD interviews — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris @NateSilver538 (or a day's worth of work to create an overfitted model that will break in the next election) — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris @NateSilver538 i totally buy that--and that's why our demographic analysis/prior MRP efforts exclusively use telephone survey data. but the same thing holds: you can't convince your model that clinton's really only at 22 percent among GA whites or whatever without...a state poll's worth of data — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris @NateSilver538 and a lot of that work is being done with state polls in the way other nate is describing earlier — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris @NateSilver538 yes, though there's a lot going into ironing out those biases on a plausibly predictive basis that goes well beyond the standard MRP stuff (though even the most heroic effort isn't going to get Indiana as a tossup in 08, or something) — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris @NateSilver538 similarly, you now keep tweeting estimates with biden doing better in georgia than pennsylvania. that's not a coincidence — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris @NateSilver538 yes, but for the wrong reason: regularization systematically diminishes Democratic strength among white voters in the north with respect to the raw data, and the opposite in the south. In '16 that makes you look smart, but not for the right reasons — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris @NateSilver538 (and it just isn't true that the principle difference between pollsters is the joint distribution of census demographics or their LV screen. plausibly true within frame/model, like among national RDD polls. probably not otherwise) — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris @NateSilver538 (as an aside, he's right that mrp means you only get one pollster's data, perhaps sometimes for the better but often for worse, unless you have microdata from multiple polls. don't really see a way around that argument, whatever the other merits) — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris @NateSilver538 mrp is less biased than the polls in '16 in no small part bc regularization means you often underestimate dem strength in the midwest and overestimate it in the south, which why you get gelman predicting clinton wins MS and why you keep showing biden up 6 in GA — PolitiTweet.org

Nate Cohn @Nate_Cohn

RT @kevinroose: The #2 most-engaged post on Facebook today is a Breitbart video of a group of doctors claiming that hydroxychloroquine is "… — PolitiTweet.org

Nate Cohn @Nate_Cohn

RT @tackettdc: "There was one problem: Mr. Trump had not actually been invited on that day by the Yankees, according to one person with kno… — PolitiTweet.org

Nate Cohn @Nate_Cohn

RT @oliverdarcy: News: Sinclair tells me it is totally abandoning the segment pushing a Fauci conspiracy theory. Previously, the company sa… — PolitiTweet.org

Nate Cohn @Nate_Cohn

@LoganDobson that's why i said facially, though perhaps not clear — PolitiTweet.org

Nate Cohn @Nate_Cohn

At least random deletion is facially unbiased https://t.co/IFvld9YiL3 — PolitiTweet.org

Nate Cohn @Nate_Cohn

RT @AshleyKirzinger: LATEST: KFF July Tracking Poll finds voters' views of how President Trump is handling key issues dropped since May 202… — PolitiTweet.org