re: Why Nate Bronze is full of shite: Using his own data

I think what is going on here is that poll have not caught up with reality.

People don't answer poll calls unless they are highly motivated. Trump voters knowing what happened in the past are not giving those people the time of day. I am not anyway.

Old school polling is obsolete. Hopefully this election brings that to light.

quote:

---

Trump is going to win Florida, Georgia, North Carolina, Arizona, Iowa, and Ohio.

Give all those and what is Trump's %?

---

before i run it, i'm going to say 80

you're closer. 63%

*ETA: if you ONLY give PA, the model goes to 65% Trump

This is beautiful

His whole 538 empire is built on pillars of salt and pillars of sans

I’ve said it before. Election Day 2016, he had Hillary winning FL, NC, PA, MI, WI, Iowa and AZ. His track record on swing states is abysmal.

quote:

---

His model was reverse engineered and some oddities were found. Good write-up here:

---

The statistician who wrote this, Andrew Gelman, is highly respected and well-renowned. He analyzed a ton of studies, and really tears apart statistically those ridiculous studies that find some asinine significant relationship and grabs all the headlines (like the study that found a relationship between the masculinity/femininity of hurricane names and the damage, deaths, etc. of it).

So when it comes to statistical modeling Gelman >>>> Nate (and most everyone else for that matter).

Nate has been in a bit of a feud/competition with the the creators of The Economist’s forecast model (specifically G. Elliot Morris) who gives Biden a 95% chance of winning (has been as high as 97%). And Nate has argued that it’s been TOO confident in its forecast of a Biden victory.

What is interesting, and I just discovered this, is one of the individuals who helped create the model is Andrew Gelman. So the foremost expert statistician, who is dissecting Nate’s model, is behind a model that believes Biden is nearly 2.5 times more likely to win than Nate’s model.

quote:

---

before i run it, i'm going to say 80

you're closer. 63%

*ETA: if you ONLY give PA, the model goes to 65% Trump

---

This makes complete sense. Pennsylvania was one of Trump’s closest victories, and is correlated with all of those states, and all of those states had larger margins, and the polling is more favorable to Trump in all of those states in 2020. So if he wins Pennsylvania, his chances to win all of those states, which are already higher, increase.

Most importantly, if he wins Pennsylvania again, Michigan and Wisconsin also become more likely, which were all super close and highly correlated with Pennsylvania. And Biden’s best path to victory is winning back those 3 states with Pennsylvania being the closest, not only because of how close they were in 2016 and how favorable the polling has been in 2020, but because they are highly correlated.

quote:

---

Aren't these models dependent on accurate polling?

---

Mostly, but it does include fundamentals as a prior, which are actually favorable to Trump (stock market, 3Q GDP) based on historical data; however, given the uniqueness of the pandemic, these data may not be as applicable as they normally would be.

That being said, the model assigns a high level of uncertainty to the polling data, using a fat-tailed distribution, based on historical polling errors (which have gone both ways like in 2012 vs. 2016 and despite polling being more accurate in recent elections), with added uncertainty due to the changes in election laws (more mail-in voting and early voting).

So by my tough calculation, the model has about a 5% standard error, which is both larger than the historical errors and the 2016 miss.

On top of that, polling has improved significantly addressing 2 problems from 2016: (1) not enough state-level polling which increases error; and (2) adjusting for education which is why Trump surprised in 2016 since this education gap emerged (specifically among whites) and those without a college degree get under-sampled. This adjustment alone probably accounts for about a 1-2 point swing towards Trump (depending on state).

I’m addition, in 2016 there was much more movement in the polls throughout (large swings), far more undecideds, and far more 3rd party supporters, all of which added uncertainty and appeared to favor Trump (Comey letter, undecideds breaking for him, 3rd party support breaking for him).

So altogether Biden’s margin is about 5-6 points larger than Hillary’s, with polling adjustments that make the polls AT LEAST 1-2 points closer (so at least 6-8 points larger margin using 2016 polling adjustments), with fewer undecideds and 3rd party supporters to tighten the polls, and more consistency in the polling. Frankly, if anything, I think there is a better chance that the polling will miss in Biden’s direction, and Nate’s model is actually overly conservative.

That’s cherry picking. His overall body of work in 2018 was quite good, 96% on 500+ elections.

I don’t understand the hate for 538. Their body of work overall is as good as anyone’s if not better. 2016 doesn’t change that.

Nate Nickel is the man!