Thread by @gelliottmorris

I think there is still a fairly sizable chance of a systematic error in the polls, and how big it is depends on which forecasts/averages you look at. (This is an important thread, please read it carefully.) 1/8

2/8 While more battleground-state pollsters are weighting by education than at this point in 2016, about half still aren't. And there are other issues with pollsters who have clearly politically biased samples because of other errors (like weighting to the 2016 exit polls).

3/8 We adjust for some of this bias by including a term in our model that balances systematic difference between correctly and poorly-weighted polls. In 2016, it shaved about 2pts off Clinton's margin in the Midwest, so it IS helpful. BUT the model was still surprised on Nov 8.

4/8 Right now, this adjustment is currently shaving about 0.5 points off of Biden's average (more in some states). We should expect, on balance, that other forecasters/averages/etc who don't adjust for these biases will overstate Biden's margin by roughly the same amount.

5/8 I do wish that the other leading forecasters would take weights and polling design into account in their models. We know that not all pollsters act responsibly and we shouldn't assume that simply averaging data will remove the biases that injects into our models (see: 2016).

6/8 At the same time, we might be able to make improvements over how our model currently handles these factors. Our bias correction is formally defined to capture the difference between pollsters who weight by party reg/past vote/etc and those that don't.

7/8 This indirectly captures SOME of the bias from the education-weighting problem, but not all of it. In future versions of our model we might be able to further refine the 2016 and 2020 bias corrections by adding another adjustment explicitly for education-weighting.

8/8 I should clarify that all these adjustments are only factored in to our _average_ projection of election-day vote shares. Our margin of error is separately calibrated to capture historical polling errors, & the final 9pt MOE on state averages would be the same no matter what.

Let me help clarify: Our model still simulates a wide range of universes of polling error, including many (about half!) where polls actually underestimate Biden. But that's after we adjust for (some of) the errors introduced by not weighting by edu. 1/2

2/2 If a model ISN'T adjusting for bias from some polls not weighting by edu or party, then _on average_ we should expect them to overrate Biden's chances. But there is still a distinct chance of some new variable or phenomenon causing bias in a diff way, hence the uncertainty.

So, what I'm DEFINITELY NOT saying is that all polls and averages are rigged toward Biden and you should ignore them. What I AM saying is that there are reasons to believe that Trump beating the polls is likelier than Biden beating the polls, even though both are still possible.