Sorry to double post but the 538 bit above got me thinking about something beyond the obvious bit regarding battleground states’ partisan leans and how they are and are not shifting per current poll results …
The perennial bugaboo of likely voter models on forecasts. There are variations of the secret sauce but generally they rely on some formula of stated intent to vote, whether or not they voted last time, and weighting by demographic past performance.
Let’s focus on Pennsylvania for an example.
Pennsylvania leaned slightly D in 2012. In 2016 it leaned slightly R. 538’s poll aggregation going into the 2016 election had Clinton up by 3.7. But we all know what happened: some who voted D in 2012 and said Clinton did not show up on Election Day, even though polls counted them as “likely” to, and some more who said Trump showed up, even though their lack of voting last time had them not pegged as “likely” to.
It’s probably not the only reason for individual state polls to have been off by a couple of percent, but it seems that LV models missing the mark with shifts in states relative to their previous leans was a big part of the state polling misses.
Was 2016 an aberration? If so do atypical turnouts last time (and thereby being a past voter) accurately predict LV behavior this time or not?
Atypical situations trigger atypical turnouts which deliver atypical = unpredictable results. Two boring establishment candidates such as Clinton vs Dole or Bush 2 v Gore probably run pretty true to historical form.
The novelty of Obama in 2008 and to a lesser degree in 2012 or both Trump and Hillary in 2016 knocked the pollsters for a loop. based mostly on the people who did or didn’t turn out when the opposite was expected.
Clearly 2020 will be hugely atypical. Between Trump, COVID, BLM, the RW militias, the USPS, there’s almost nothing about this election that is ordinary. Except maybe Joe Biden. Although even he sauced that up (for good or ill) with Kamala Harris.
The other gotcha is that as turnout has been on a downtrend for decades, the voting power of the normally-apathetic set is amplified. It only takes something motivating a smaller and smaller percentage of that larger & larger cohort to upset the predictions.
For all these reasons I personally put very large error bars on top of the pollsters’ error bars.
I’m nervously completely agreeing that the extreme atypical nature of this specific time will make LV models very very hard to do well, and as you put it superimposing extra large error bars is indicated.
But I think you are mistaken about that “other gotcha”. See here.
From 1972 to 2000 presidential election voter turnout averaged mid 50%s. In 2004 it hit over 60% for the first time since 1968 and only dropped slightly below it in 2012, staying over otherwise. 2016 was 60.1 and only 2008 at 61.6 was higher since 1972 (2004 tied)
Midterm turnout also pretty flat low since 1974, 38.1 to 42 range, dropping in 2014 to a miserable 36.7. But in 2018 jumped to 50%.
There is really no long term downward trend of turnout.
But does huge midterm turnout in the 2018 presage big increases in 2020, or do various obstacles, from pandemic to voter suppression to just lack of traditional GOTV approaches, tamp it to lower than 2016 or even 2012 numbers? And who is actually likely to vote despite those obstacles?
Texas is a pretty conservative state. Any time I see references to Texas’s high Latino population, I always wonder if Texas Latinos are more conservatives than Latino Americans in general. I skimmed the report you linked to and didn’t see that info there.
I was looking at Wisconsin results a couple of days ago. At the time, the best pollsters were rated B/C, and Biden was ahead by 4.3% (across six polls), with Trump leading in one (by one) and tied (in one). Biden was ahead by 10% in one poll (that has to be an outlier).
Yes. If Trump loses FL (and assuming PA is safely blue), he needs to run the Midwest table by winning OH, MI, WI and MN. That’s not impossible, but it’s really unlikely (he lost MN in 2016, and nobody thinks he has a chance in hell in MI).
This may have gotten lost in the big batch of polls I posted Tuesday, but Morning Consult did a neat little experiment here. They polled each of 11 swing states both immediately before and immediately after the convention in an effort to measure convention bounce. It’s nice to get polls from the same pollster, with the same methodology, in the same timeframe, across a bunch of states.
Here are those polls again, grouped nicely for easier reading. This shows a slight shift toward Biden on average, but most of that is a huge shift in Arizona. Without Arizona it’s basically flat.
Which is why I always argue the Ds should avoid the HRC mistake of 2016 by putting all their eggs in the Florida basket while ignoring other, less flakey states.
Newest polls in 538 show improvement for Biden in North Carolina, Minnesota, and Arizona — enough put him back at winning 70 out of 100 iterations of their model.
Yep, they’ve changed it from Biden is slightly favored to win to Biden is favored to win. I’m guessing those new polls have nudged things in that direction.
)