Nate Cohn

Nate Cohn @Nate_Cohn

ABC/Post comes in at Biden+12 among RVs, 53 to 41, with a 10 point lead among LVs. That's Biden's best result among live interview polls of the month, and it has the same, late 8/12-15 field period as CNN/SSRS — PolitiTweet.org

Nate Cohn @Nate_Cohn

There's a fun debate about how to think about these kind of things. One could argue you want to simulate the ballot; another argument is that most voters make up their mind before actually voting, so simulating the ballot may emphasize things that don't actually influence voters — PolitiTweet.org

Nate Cohn @Nate_Cohn

One interesting thing is that the horserace question lists the full ticket, which isn't unheard of but also isn't typical either. I'd be curious to know whether there's an effect there https://t.co/jEdWkw0J6h — PolitiTweet.org

Nate Cohn @Nate_Cohn

CNN poll shows Biden+4, 50 to 46. Obviously one of the best national survey's Trump has had in a long time, and with a few interesting notes under the hood — PolitiTweet.org

Nate Cohn @Nate_Cohn

RT @rickklein: @POLITICO_Steve We had a first taste this morning. Full results tomorrow. We went in the field after the Harris pick. — PolitiTweet.org

Nate Cohn @Nate_Cohn

RT @PatrickRuffini: Just a quarter of Latinos have even heard of the term “Latinx” and only 3% use it. https://t.co/gPs0uQRq5E — PolitiTweet.org

Nate Cohn @Nate_Cohn

RT @ForecasterEnten: The USPS is the most liked federal agency... and... is equally liked by Dems and Republicans. https://t.co/6GeQymjCEJ… — PolitiTweet.org

Nate Cohn @Nate_Cohn

RT @AlxThomp: NEW: Inside “Uncle Joe’s” mission to prove Obama wrong & the complicated relationship Biden “oftentimes felt that that loyal… — PolitiTweet.org

Nate Cohn @Nate_Cohn

@NateSilver538 @gelliottmorris @DanRosenheck @SethS_D whether and how to consider presidential approval seems relevant here — PolitiTweet.org

Nate Cohn @Nate_Cohn

RT @AnnaGHughes: Many people have been using quarantine as a time to perfect their bread or coffee making skills, but I personally have tak… — PolitiTweet.org

Nate Cohn @Nate_Cohn

There are many reasons why we'd expect this: language barriers, political engagement correlating with interest in responding to a poll, etc. — PolitiTweet.org

Nate Cohn @Nate_Cohn

RDD pollsters, like Marist, take a sample of adults and then screen to RVs based on self-reported registration. If your respondents are likelier to be RVs than the actual population, your screen will not create a large enough RV-adult gap https://t.co/liHos3VgpA — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris and if i ever do a forecasting model, it would probably be an ensemble of various choices--probably inspired by my initial weather interest--rather than commit to one and pretend that my choices are certain — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris when you have models for this purpose, you don't really commit to one model! i had one with or without polarization, to take the example here. that's kind of useful if you want to know whether a CD like NC09 is competitive under certain circumstances. https://t.co/roF73CwVfO — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris there are times where i do it: i had a house model in '18. it's really hard to think about house control without one, given how many seats there are, the sparse polling, etc. for a presidential race? not a ton of value add over a polling average and common sense, imo — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris if folks want to go and do it, i'm not saying the world is a worse place for it. but it also means there's even less value if i go and do it too. — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris a robust forecast at N=17 means making up a bunch of stuff, and it's a lot of work to inevitably say 'candidate who leads the polls will probably but not necessarily win,' and then be yelled at if the polls are wrong, as they will be from time to time. i don't see the fun — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris fair enough, but we're just not anywhere near the same page on how to think about this stuff if there's any utterance of the word 'underfitting' over the last... 5 data points ? — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris instead, i see concern about underfitting? what? like, a good model probably should be underfit at N=17! why would you assume you've seen anything near the full range of possibilities — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris for me, this is the kind of uncertainty that matters most. maybe it's because i read history a lot. maybe it's because i was interested in weather before politics, and see what you can't get here. but none of your tweets or methods seem focused on it. — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris it does not seem to me that this kind of uncertainty figures in your thinking. it's not going to come up with out of sample cross validation on N=17. regularization doesn't fix it. but it's there and it's big at N=17 — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris i look at this chart and see an example of a model that would have worked pretty well, and now doesn't. to me, it shows the limits of empirical forecasting. things change. you look and it and think you have it solved with a more complex model. — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris one nice benchmark is the number of variables considered with respect to the sample size, and you're talking about an N=17 or whatever dataset! In the absence of strong theoretical principles, you don't have the data to make durable out-of-sample inferences with a complex model — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris As someone who likes MRP, surely--surely--you have experience fitting a model, adding more and more and more variables without meaningful change in fit, but nonetheless get more and more outlandlish predictions on unusual cases — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris I don't see what this purports to prove? Just because a model doesn't have a really tight fit doesn't mean it isn't overfit? It's not like you're overfit at +/- 4 but not at 6 or something. It just doesn't really help evaluate this stuff — PolitiTweet.org

Nate Cohn @Nate_Cohn

RT @theJackVaughan: Election Twitter’s crowdfunded KANSAS poll results SurveyUSA/Election Twitter (8/5-8/9) MoE +/- 3.3% ________ KS-SEN… — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris i mean, looking at this to evaluate whether you're overfitting or not is kind of getting at the central area of disagreement in how to think about this. but also, don't these look... really overconfident? that dotted line sure falls outside or on your 90% CI a lot — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris Maybe it was too tough of a word choice. I'm not trying to be harsh here; he asked where the difference was, and I think I've outlined just a really core difference in how to think about uncertainty in model design https://t.co/vOuPGgAgsb — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris Here's maybe the normie way to look at it: when you keep tweeting about how well some choice validates, and constantly emphasize the *statistical* tools you choose to avoid overfitting, not the analytical ones, that scares me and indicates that you're falling into a trap — PolitiTweet.org

Nate Cohn @Nate_Cohn

@gelliottmorris I'm not trying to be harsh: these are tools that are intended to reduce the danger of overfitting, and you can really trick yourself into thinking that means you're not overfitting. https://t.co/qiZ6nYDvCW — PolitiTweet.org