Thread by @gelliottmorris

I've gotten a lot of questions about our model's projection in Arizona today (probably because of the NYT poll). People think it should be bluer. Maybe! Our model has it closer to 50-50 (though on the verge of lean blue) than other models (eg 538) for a few reasons.

https://twitter.com/gelliottmorris/status/1313193214820708372

First, our model is picking up some real pro-Biden pollster-level and mode effects in many of the polls coming out of AZ. I remove them, Biden's current adjusted polling margin there increases from three to four percentage points, and same with the election-day prediction.

A second factor is that polls were a lot closer a week or two ago, and our model thinks that information is still valuable. And since Biden has only gained ~1 point or so nationally and in states like AZ over that period, the model doesn't want to adjust too much to newer data.

(As a brief aside: our thinking on polarization also seeps into how our model calculates trends. Bc we specify a prior MOE of about 20pts of movement in any state over the course of the election cycle, it's primed to think that any 1 or 2wk shifts should rarely be very large.)

(So if we had as much uncertainty in our model as some of the other forecasters, trends at the state level might be a bit more volatile, and AZ and OH would be bluer right now.)

Finally, our model factors in data other than the polls: Namely, the implied result based on each state's predicted partisan lean and the current national margin. That's closer to Biden +2 than the ~+6-9 that some polls have been showing.

One caveat here: As is also potentially the case in Texas right now, that prior can be thrown off by a surge in new registrants or other state-specific factors between this election and the last (also see Iowa 2012-2016). So there is quite a large margin of error on that prior.

Still, the best practice historically has been to combine polling data with other sources of information in making our predictions. And while the model will be ~80-90% poll-based by election day, until then we should continue to be relatively bearish on AZ projects.economist.com/us-2020-foreca…