Thread by @gelliottmorris

I wonder what the @FiveThirtyEight and RealClearPolitics averages would show if they excluded polls released by Trump's super PAC (Insider Advantage, Susquehanna) and other ideologically skewed firms with suspect methods (Rasmussen, Trafalgar). This is actually a really big deal.

There is good reason to believe that the way house effects and partisan bias adjustments have been calculated in the past — to control for wonky methodologies or campaign internals — aren't enough to control for the several GOP firms that are essentially catering to the right.

There is also a point about incentives here, which is that the polls that Trump's Super PAC choses to release are probably intended to raise money and motivate his base, rather than accurately measure the reason. That's even worse than the raw Ras + Trafalgar bias

In the swing states like MI, PA and FL with a ton of these polls, all this super-GOP-leaning data is probably creating a roughly ~1 bias toward Trump on vote margin, with a slightly large probabilistic shift. That explains a lot of the delta between our model and the others.

It might be obvious by now, but there are a lot of incentives for forecasters post-2016 to WANT to include data that let us implicitly hedge by pushing our models back toward 50-50. But there's a huge data integrity tradeoff here, and I'd rather not pollute my models w crap data.

Another point is that there AREN'T similar pollsters like these on the D side, so the idea that biases will cancel out is silly.

But again, the bigger deal is w house fx, which are 1-2pts for Susq/Insider/Traf/Ras in 538's model but should be more like 6.

https://twitter.com/tehoriman/status/1321820807333556225?s=20

(The only dem-leaning pollster with a remotely similar house effect is Y2 Analytics, but that's because they didn't weight by education in a couple of Utah polls earlier this year and also had a really outlier aberrant poll. Nothing with the frequency of these R firms.)