Thread by @gelliottmorris

Here's my final update to this model of California recall polls. I'm calculating an aggregate that adjusts polls based on whether they use partisanship in their weighting schemes, and draws different trends for adjusted v unadjusted data. Newsom +18 +/- 10
gist.github.com/elliottmorris/…

https://twitter.com/gelliottmorris/status/1435699656202280967

The point of this project was to illustrate the different methods we can use to aggregate polling data — esp in how to improve existing popular averages that don't peer under the hood of how pollsters are processing their data, an increasingly important aspect of public polling.

So, note two things:

1. Popular averages magnify unlikely trends in public opinion by being whipsawed by data that is subject to higher standard errors than a decade ago (when the models were first made). Weighting by party flattens trends by decreasing nonresponse

2. The uncertainty of the model increases as we adjust for more factors. The uncertainty may seem so large as to make the aggregate useless, but that's a little silly. Instead, I think these intervals better represent the total survey error in polls — that 538 etc have ignored.

Will probably blog more about this modeling later. I think it's important and represents an innovation over decades-old averaging techniques w/ ad-hoc adjustments for past accuracy that don't capture what's actually happening in the data-generating process gelliottmorris.substack.com