USC/Dornsife tracking poll megathread

[tagimaucia]

https://twitter.com/Nate_Cohn/status/1296085937642971136

Those numbers are eye-opening stuff.

There were something like 14 million voters in the 2018 midterms who didn't vote at all in 2016, IIRC.  That number will obviously be even higher in 2020 (relative to 2016).  That's why Trump can't just cling on to his 2016 voters IMO.  He has to expand his standing with people who *did* vote in 2016 at least a bit to have any shot at winning.

[tagimaucia]

Another Trash Poll, also not included in the RCP average
 

Is this a bit? RCP is actually known for having no clear criteria or transparency for what they include in their averages and they include every poll by many pollsters universally known to be crap while completely blacklisting or cherry picking many other ones with good track records.

[tagimaucia]

Even though this poll was off by ~5 points in 2016, it showed the movement towards Trump in the week or so before the election a lot more clearly and distinctly than most polling aggregates did, and also showed that it continued to build basically until Election Day.  I think its worth at least keeping an eye on, just add a handicap of however many points if you think its too friendly to Biden.  And its' too soon to say its not showing any RNC bounce IMO. 5/7 of the current sample is still before the final night of the convention, although at this point it doesn't seem likely that it will show a big one.

[tagimaucia]

Just found this.  Back in 2016, this economist constantly reweighted the 2016 LAT/USC poll based on the American Community Survey and other demographic data.  Not only did his "reweighted" version come within 0.2% of the final margin on Election Day 2018 (his final margin was Hillary +1.9, the actual was Hillary +2.1), but his estimations basically worked perfectly as a leading indicator of the RCP + other poll aggregates that was ahead of them by a week or a little less.

https://twitter.com/ernietedeschi/status/796053756915908608

https://sites.google.com/site/latuscrw/

[tagimaucia]

I think the stuff I posted earlier shows that this pollster is very good and useful for accurately showing movement in the race-- they just really really sucked at weighting their results properly in 2016.  They did a good job in 2012 (apparently the RAND American Life panel from that year was run by the same people) so hopefully they've fixed whatever was wrong last time, but who knows.

[tagimaucia]

So this poll shows the DNC and RNC bounces as basically exactly equal and offsetting (~4-5% for each), and the horserace at pretty much right where it was before the conventions.

[tagimaucia]

.

[tagimaucia]

I think it shows Hillary’s steepest decline actually happening just before the Comey letter, which lines up with this analysis: https://www.nytimes.com/2017/05/08/upshot/a-2016-review-theres-reason-to-be-skeptical-of-a-comey-effect.html

[tagimaucia]

Sigh. This could be a valuable poll if it was run better but the people in charge of it seem to be incapable of not making stupid methodological decisions from year to year. The oscillating nature should have maybe been obvious but at first it seem to line up perfectly with the conventions and where you would expect bounces to be so it fooled me. Lol.

[tagimaucia]

Probabilistic:

8/22:  Biden +10.8
9/5:    Biden +7.8
9/19:  Biden +8.7

¯\_(ツ)_/¯

[tagimaucia]

So I guess we’ll be in the pro-Biden half of the sample on election day, right? (6 weeks plus 2 days)

Hopefully they'll at least have the decency to cram everyone in the panel into their final 7 day sample before the election.

[tagimaucia]

LOL.

[tagimaucia]

https://twitter.com/ernietedeschi/status/1313099199622320130

[tagimaucia]

Here's what this poll looks like when you take the 7-day tracker, make bias adjustments for each little slice of the 14-day "cycle," and then add some smoothing.

[tagimaucia]

Here's what this poll looks like when you take the 7-day tracker, make bias adjustments for each little slice of the 14-day "cycle," and then add some smoothing.

Updated this again because why not:

[tagimaucia]

The 7-day average is actually surging right now, to a degree that I don't quite believe it + think there might be some kind of data problem again like the one they fixed a few weeks ago.  Look at my graph or compare the 14-day results today to those from 14 / 28 / 42 / etc. days ago and you'll see what I mean.

[tagimaucia]

I hope so!

[tagimaucia]

Final 14-day track:

Probabilistic:

Biden 53.29 (+10.49)
Trump 42.8

Traditional:

Biden 54.03 (+11.79)
Trump 42.24


My best estimate of the (adjusted for cohort) final 7-day track:

Probabilistic:

Biden 53.17 (+10.07)
Trump 53.11

Traditional:

Biden 54.46 (+12.17)
Trump 42.29