There are plenty of places where the polls have been 'wrong' in recent cycles.
One place where they have not been wrong, in either 2016 or 2018, is Texas.
Now, there's no reason to assume that the distribution of poll error and bias will be the same in 2020 as it was in 2016/2018.
But that assumption is *usually* one of the better reasons to think Trump might be in better shape than polls suggest.
With Texas, we have a state where basically every poll for the last year shows a total tossup, where there's no history of systematic polling error during the Trump era, and where it is awfully hard to believe that the polls could be right without Biden being a heavy fav
So if you're in the 'polls surely biased toward Biden bc 16/18' camp, I do think it's worth pausing at these Texas numbers and considering the extent that they complicate your theory of the case.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
One thing I've been experimenting with since our GOP oversample this summer: weighting our polls by each partisanship subgroup, which has the consequence of ensuring that each subgroup used in weighting has the right number of Dems and Republicans
The main downside is that our estimates for self-reported education by voter file party are modeled, and that's something I've had pause about. I'm gradually getting more comfortable with it, as the party x edu tallies for the typically weighted sample seem consistent with subgroup
Perhaps surprisingly, there's essentially no difference in our topline results between polls weighted by party and those weighted across the full population. The differences by subgroup are surprisingly minor, as well
I happened to be looking a lot at Pew data last few weeks, even before this most recent partisanship study, so I wanted to share a few interesting observations about trends I noticed in their data pewresearch.org/politics/2024/…
One thing I noticed: subtle but persistent, multi-year differences between the partisan splits by demographic on the Pew ATP -- the mostly mail-to-web panel they use for this study -- and the Pew NPORS study (the one-off high-incentie mail survey they use for weighting the ATP
The most striking thing, IMO, is that the Pew ATP consistently has more age and racial polarization than the NPORS
A few outtakes:
-- By our (rough and preliminary) estimates, this looks to be yet another zero-persuasion (off Biden '20) special. We'll have to see final vote history, but at least in Nassau it looks just as we'd expect given the party reg turnout
-- We'll see how the dust settles, but I do think it could be significant if the result is interpreted as showing Dem strength on the border messaging. If that narrative takes hold + Dems are emboldened to follow on, that's quite helpful on their worst issue
1) Specials are driven by turnout. The data is unequivocal as long as I've been looking at them with our rich data, going back to 2017. That should not be remotely surprising, as the as the people who know about/vote in specials are highly parstisan -- just like all of you!
2) Special elections therefore tell us something about enthusiasm/engagement among highly engaged voters, which is helpful in predicting low turnout contests. It is less significant as turnout increases, but may still helpful to Dems (see favorable LV/RV gap in NYT/Siena '23)
We have a new NYT/Siena national survey, and it's an interesting one -- with the public sympathetic toward Israel but disapproving of Biden on the issue and split on whether Israel should continue military operations.
It also has interesting '24 numbers... nytimes.com/2023/12/19/ups…
Trump leads 46-44 among RVs, but *Biden* actually leads in our first measure of the likely electorate nationwide, 47-45.
The split is driven by a huge gap in vote choice by turnout history: Biden+6 among '20 voters; Trump+22 among 2020 nonvoters nytimes.com/2023/12/19/ups…
We've seen a similar pattern in our polling over the last year, with Biden excelling among regular and esp '22 voters.
But this is the largest split by '20 vote, and as a consequence it's a bigger LV gain for Biden than prior polls would have shown.
One thing I've seen over the last few days: a lot of people asserting that, in a variety of different ways, pre-election polls aren't very useful for demographic subgroups
I have to completely disagree.
Stepping back, it's my long-standing view that the pre-election polls are the best basis for post-election estimates -- and, in particular, better than exits.
For ex, all these estimates were based on pre-election polling: nytimes.com/2016/06/10/ups…
This was already true back in 2012 and 2016, but it's become indisputable in the era of early voting. The exits are mostly pre-election polls at this point. AP/Votecast is just a pre-election poll. Catalist is also derived from pre-election polling. And so on.