How do pollsters model for “likely voters”? That question takes on larger importance this year, as the COVID-19 pandemic, riots, and widely divergent campaign styles of the two presidential contenders raise questions about turnout impact. Polls based on responses of “likely voters” could be less predictive than in previous cycles if pollsters can’t figure out how to factor all of those issues into their models.
Or perhaps this year won’t be any different, as Mark Mellman argued yesterday at The Hill. That’s not because pollsters can handle the adaptations needed, the Democratic pollster writes. It’s because likely-voter screens and models are a “sham” already:
Last year, political scientists Anthony Rentsch, Brian Schaffner and Justin Gross analyzed the 64,600 interviews from the 2016 Cooperative Congressional Election study and found just 64 percent of those who maintained they were “definitely” going to vote did so. Only 68 percent of those who claimed they had already voted actually cast a ballot at any point. Meanwhile, 18 percent of self-reported “less likely” did turn out, as did 9 percent of those who claimed they had no intention of voting.
That question may be particularly blunt, but more complex items haven’t proved superior.
Among those who scored the highest on Gallup’s multi-item scale in 2014, 83 percent voted, but 17 percent didn’t. Moreover, more than 20 percent of those who did cast ballots scored at the bottom end of the scale and would have been removed from the “likely voter” pool.
What about surrogates like “enthusiasm?”
In one race, we found among those “very” enthusiastic, 88 percent voted; among those who were “somewhat” enthusiastic, 83 percent turned out; and among those “not too” enthusiastic, a slightly larger 85 percent turned out. No relationship.
Moreover, the “unlikely” voters that get overlooked actually tend to vote after all, Mellman states. That happens at roughly the inverse rate of turnout among identified LVs, he says; 80% of LVs turn out, while 20% of UVs end up casting ballots. That’s enough to make likely-voter models non-predictive, Mellman argues, stating that Harry Reid’s surprise win in his last election bid resulted from a reliance among analysts on LV polling.
So how should people view election potential? Mellman looks for the “likely electorate,” a concept that he doesn’t quite identify:
First, our methods for determining who is a likely voter aren’t very effective (particularly if you aren’t using a voter file). Second, we aren’t really interested in likely voters. We should rather be interested in the likely electorate and the two are not the same.
That’s great, but … what is a “likely electorate”? It’s a mix of many likely and some unlikely voters, Mellman writes, but that’s not terribly helpful from a data standpoint either. Which members of each group should we include, and which should we exclude? That question is particularly acute in 2020, where more stresses on voters exist that could impact turnout behavior.
It seems that the lesson here is not to rely on modeling at all, but on other data to determine electorate behavior. In 2020, that would leave us with voter-enthusiasm measures, especially in personal contacts by campaigns and secondary indicators such as swag, yard signs, bumper stickers, and so on. Allahpundit already wrote about the NYT analysis of rising Democratic panic about Joe Biden’s lack of a ground campaign, but it’s worth bringing up here in the context of Democrats’ attempts to take comfort in the polling:
“It feels like asymmetric warfare,” said Matt Munsey, the Democratic chair in Northampton County in eastern Pennsylvania, one of the counties Mr. Trump narrowly flipped in 2016, referring to Mr. Biden’s approach versus Mr. Trump’s.
Livestreamed events were “not necessarily reaching people,” Mr. Munsey cautioned. Mr. Biden has begun to accelerate the pace of his travel, and Mr. Munsey praised him for “getting out there more.” But he expressed frustration that Mr. Biden’s in-person events were kept so small: The campaign has been so wary about exceeding crowd limits, he said, that local leaders have complained of not being invited.
Compounding the challenge is an on-the-ground operation that was weak during the primary season and was slow to scale up in the general election. Strapped for cash after the primaries and uncertain about how to campaign amid a national lockdown, the Biden team initially refrained from greatly expanding its staff. It entered the summer without state directors in critical battlegrounds like Michigan, Florida and Pennsylvania, and efforts to establish local operations stretched deep into the summer.
Now Democrats from Florida to Nevada have worried that the team is behind where it should be in engaging some core constituencies, a problem that may also have implications for new voter registrations.
“Asymmetric warfare” assumes both sides are engaged. In the ground game, there is no warfare at all. Trump’s getting a walkover on it, and it has produced some interesting and measurable results outside of polling. In Pennsylvania, for instance, Republicans have added new voter registrations at a 4:1 or 5:1 clip over Democrats since 2016. It seems likely that similar numbers exist in other battleground states where the Trump campaign and the RNC have focused their ground-game efforts.
It’s admittedly tough to show a direct correlation between those indicators and outcomes, let alone construct a predictive model for that. However, Mellman’s warning is a good reminder that polling is an inexact and sometimes inaccurate look at a snapshot in time. The spate of articles focusing on Biden’s lack of person-to-person campaigning shows that more Democrats than Mellman understand that, too.
Update: Scott Rasmussen wrote about the issue earlier this week as well.
