2021 Election Post-Mortem: How much does polling drive turnout?

The final results from the 2021 elections are being counted, but the big picture is clear:

  1. Republican Glenn Youngkin defeated Democrat – and former governor – Terry McAuliffe in Virginia’s gubernatorial election.
  2. Republicans took control of the Virginia House of Delegates.
  3. Democratic New Jersey governor Phil Murphy was reelected.
  4. Democrats Eric Adams and Michelle Wu will be the next mayors of New York City and Boston, respectively.

You may find my final forecasts for the New Jersey and Virginia gubernatorial elections here. Overall, they were – fair-to-good.

**********

Let us start at the end with the safest predictions:

Democrats Eric Adams and Michele Wu are heavily favored to be the next mayors of New York City and Boston, respectively.

Both Democrats won by a roughly 2-1 margin. Adams, 61-year-old former New York City Police Captain and Brooklyn Borough President, will be the city’s second black mayor. The 36-year-old Wu, Harvard-educated daughter of Taiwanese immigrants and former Boston City Council president, will be the second woman of color to be mayor of Boston – but the first to be elected mayor. Boston City Council President Kim Janey, a black woman, became Acting Mayor of Boston when Marty Walsh was sworn in as Secretary of Labor on March 23.

Working backwards, that brings us to:

Democrat Phil Murphy will be reelected governor of New Jersey, defeating Republican Jack Ciattarelli, though the margin will be far lower than the “fundamentals” (state partisan lean, incumbency, generic ballot) suggest it “should” be.

While I was correct about Murphy winning – becoming the first Democrat to be reelected governor since Brendan Byrne in 1977[1] – I was wrong in thinking the call would be made within an hour of the polls closing at 8 pm EST. Instead, Murphy was not declared the winner until the following day. Indeed, Ciattarelli still led – albeit by just 565 votes – when MSNBC concluded its Election Day broadcasting at 3 am; I know, because I was still watching then.

The long delay resulted from two factors. One was simply the surprising closeness of the race. As of 5 pm EST on November 6, Murphy had a 2.6 percentage point (“point”) lead. However, my final WAPA (weighted-adjusted polling average) had Murphy up 7.6 points – a value which had barely changed since mid-August. The final 10 polls had him up 8.3, while a linear regression model estimated his Election Day lead closer to 7.0 points. At worst, and assuming third-party candidates earned 1.5% of the vote, if the remaining undecided votes broke 2-1 for Ciattarelli, I calculated Murphy would still win by 5.3 points. In the end, I probably took New Jersey’s strong Democratic lean too much into account, forecasting a winning margin of 8.2 points.

Oops.

The other factor, however, was simply the way in which ballots were tallied. Mail-in ballots – which tend to be overwhelmingly Democratic – could not be processed until midnight of Election Day. Thus, what was being reported through much of the evening was the Election Day vote, which tends to be strongly Republican. Both parties were reportedly extremely frustrated by the confusion surrounding the type of votes being reported and how many votes of each type remained to be counted in any given area.

My final WAPA and projection overestimated the actual margin in New Jersey by 5.0 and 5.6 points, respectively, right on the historic average. I also overestimated third-party percentage by 0.8 points. Using the actual 0.7% they earned mathematically implies the truly undecided voters broke something like 5-1 for Ciattarelli.

No exit polling was conducted in New Jersey (and Virginia exit polls did not ask when voters made their decision), so we cannot verify this empirically. Still, a 5-1 split feels implausible. A more likely explanation relates to what polling suggested about the closeness of the race. The party that wins the White House traditionally does poorly in elections over the next two years because voters of the opposite party are energized to “resist” while lower-propensity voters of the White House party do not see the need to vote again so soon and/or are disenchanted by what they perceive to be a slow pace of change. Murphy led in all 18 publicly-released polls released since mid-April 2021, by anywhere from 2 (Fabrizio, Lee & Associates; 8/24-29) to 26 points (Rutgers University; 5/21-29);[2] as I write often about Emerson College polls, I note their one poll of the race (10/16-18) showed Murphy ahead a fairly-accurate 4 points. The emerging consensus – echoed by me – that Murphy was a lock to win reelection could easily have led lower-propensity, nominally-Democratic voters not to vote.

And, in fact, overall turnout – about 2.53 million votes – was 45% lower than the 4.57 million who cast a vote in 2020 and 13% lower than the 3.05 million who did so in 2017. Moreover, Murphy is currently down an astonishing 817,907 votes from his landslide win in 2017. This dramatic decline in turnout should have doomed Murphy – yet he still won.

**********

This brings us to Virginia, where I never projected races for the state Senate or House of Delegates; still, losing control of the House of Delegates is not good news for Democrats. I focused solely on the gubernatorial election, writing:

Republican Glenn Youngkin may be poised to upset former Democratic governor Terry McAuliffe…anything from McAuliffe winning by 4 points to Youngkin winning by 3 points remains plausible, if not probable.

To be honest, despite the conditional word “may,” when I sat down around 6:30 pm Tuesday night to watch the returns on MSNBC, I still expected McAuliffe to eke out a narrow victory, based upon the final WAPA of McAuliffe+0.4 and the increasingly-Democratic lean of Virginia – and, perhaps, a bit of wishful thinking.

Instead, though, what had happened in 11 of 12 elections beginning with Republican John Dalton defeating Democrat Henry Howell in 1977 – the year after Democrat Jimmy Carter was elected president of the United States – happened again: the candidate of the party winning the White House the previous year lost the Virginia gubernatorial election. Youngkin currently leads by 2.2 points, 2.6 points lower than the final WAPA and 2.7 point lower than I projected. I take some solace in the fact the final margin was well within my suggested range of outcomes.

In retrospect, I did not give enough weight – qualitatively and quantitatively – to the sudden and rapid decline in McAuliffe’s polling – a decline I showed graphically and modeled in multiple ways (see Figures 1 and 2). This is likely because I had not closely followed the controversy surrounding McAuliffe’s September 28 debate answer about the relative roles of parents and school boards in determining what books children read in schools.

Figure 1: Linear model of McAuliffe lead in 2021

Figure 2: Polynomial model of McAuliffe lead in 2021

And I should have recalled an earlier election in a blue state during a good year for Republicans – the 2014 Massachusetts gubernatorial election – when the Democratic nominee’s polling collapsed late in the campaign. State Attorney General Martha Coakley – the Democrat who had lost a special Senate election to Scott Brown nearly five years earlier with even more egregious unforced errors – led former Harvard Pilgrim Health Care CEO and state cabinet member Charlie Baker by double-digits through mid-September. However, Baker led Coakley by an average 0.7 points over the last 23 polls, ultimately winning by 1.9 points.[3] I then supported him for reelection in 2018 in a public display of bipartisanship.

To be fair, I did observe Youngkin had taken a clear lead during the final month of the campaign, writing “the overall 10-poll rolling average went from McAuliffe+2.0 to McAuliffe+1.0 to McAuliffe-2.1 over this same four-week period” In fact, WAPA calculated using the final 15 polls of the race – beginning with a tie in the Monmouth University poll conducted October 16-19 – showed Youngkin ahead 1.8 points, very close to the actual 2.2 points (which will likely narrow a hair).

I just didn’t – accept – this, paying too much attention to the Fox News poll conducted October 24-27 showing Youngkin ahead 8 points – and a 13-point increase in margin in just two weeks. The fact that Youngkin still led by 0.9 points over this period without the Fox News poll should have given me pause. I humbly confess it did not, giving too much credence to the state of the race just before that poll was conducted:

In fact, taking just the average of the most recent Washington Post and Emerson College polls, along with one from Suffolk University…shows McAuliffe ahead 0.3 points in the raw margins and Youngkin ahead 0.2 points after adjusting for historic bias. I think these three polls give the most accurate picture of the campaign – effectively a coin flip – at least as of October 23 or so. (Boldface added for effect.)

As with New Jersey, I overestimated the third-party percentage of the vote – it was 0.7%, not 1.5%, mathematically implying truly undecided Virginia voters broke about 5-2 for Youngkin, which is somewhat plausible.

Unlike in New Jersey, however, turnout in Virginia actually increased – perhaps because the polls suggested a very close race in which the Republican nominee had a serious chance to win. About 3.28 million Virginians cast a ballot in 2021 – only 26% lower than the 4.46 million who cast a vote in 2020 and 25% higher than the 2.61 million who did so in 2017. McAuliffe is currently up fully 179,651 votes from Ralph Northam’s surprisingly-high total in 2017. Given these figures, McAuliffe should have won relatively easily – except that Youngkin is currently exceeding Ed Gillespie’s 2017 vote count by 484,696 votes! In fact, both turnout and Republican margins increased in “red” areas of the commonwealth, overwhelming better-than-expected turnout in “blue” areas of the state – albeit with smaller margins because some suburban voters who had swung Democratic since 2014 swung Republican.

**********

And that may actually be the biggest takeaway from the 2021 elections: swing voters still very much exist.

Except – for all the talk about the decline in Democratic margins in suburban counties relative to 2020 (Fairfax County, for example, dropped from D+42 to D+20), white college educated voters were NOT the swing voters. In fact, based on 2020 and 2021 CNN exit polling, these voters…

…split 52-45 Democratic in 2020 while being 33% of the electorate

…split 52-47 Democratic in 2021 while being 37% of the electorate

If anything, white college-educated women were MORE Democratic in 2021 (61-39) than in 2020 (58-41), while making up a slightly larger share of the electorate; white college-educated men became slightly more Republican in 2021 (R+3 to R+11), while making up a slightly smaller share of the electorate. That is hardly “swing voter behavior, however.

Voters of color also were a bit more Democratic in 2021 (77-22) than in 2020 (75-23), albeit while dropping from 33% to 27% of the electorate; had the 2020 demographic distribution stayed the same, McAuliffe would have won by a little under two points. Again, though, by no definition are persons of color swing voters.

No, the demographic shifting most from 2020 to 2021 is white voters without a college degree, especially white women without a college degree, who were as overwhelmingly Republican as their male counterparts in 2021, as Table 1 shows:

Table 1. Virginia white voters without a college degree in 2020 and 2021

Group20202021
 % Voters% DEM% GOP% Voters% DEM% GOP
Men173168172278
Woman174456192574
Total343862362476

Relative to Democratic presidential nominee Joe Biden in 2020, McAuliffe’s support collapsed from -12 to -49 among white women without a college degree, a truly monumental decline. Had this group simply split the way white men without a college degree did in 2020 (31-68) – when Republican presidential nominee Donald Trump was on the ballot – McAuliffe would have eked out a 5,600-vote victory (~0.2%); had they voted as they did for Trump one year earlier, meanwhile, McAuliffe would have won by nearly 5 points.

That is what swing voters look like.

I do not know why non-college-educated voters became even more Republican in a year when Trump was NOT on the ballot, but – to me – this is far more worrisome for Democrats than a typical decline in the non-white share of the electorate and modest declines in Democratic margins in suburban areas; both will shift back towards Democrats in 2022, especially if Republicans nominate far more Trump-like candidates. As I wrote elsewhere, Democrats do not have a turnout problem, they have a Republican turnout problem.

**********

All of which brings us back to public polling, from which I calculate WAPA.

I surmised that after consistently undercounting Republican percentages in recent elections, pollsters had overcorrected, and were now overcounting these percentages.  That is clearly not true, however: my final WAPA missed the actual Democratic-minus-Republican margin in the New Jersey and Virginia gubernatorial elections by an average of -3.8 points. In 2020, state-level presidential voting WAPA missed the final Biden-Trump by an average of -3.4 points, while missing Democratic-minus-Republican margins in Senate elections by -5.8 points; the average miss in gubernatorial elections were skewed by two extremely large misses. WAPA similarly overestimated Democratic margins in 2018.

There has traditionally been no pattern to the partisan direction of the average polling miss in an election year – but pollsters have been primarily overestimating Democratic strength since the 2016 elections. I have thus decided I am going to automatically deduct 4 points from all WAPA, and use THAT value for projecting final margins – at least until pollsters determine how best to count Republican voters. I also need to figure out how to account for sharp, late shifts in WAPA – though this adjustment may already do that.

One immediate consequence of this automatic deduction: I currently have Democrats leading in the generic ballot polls by 2.3 points. Deduct 4 points, though, and Republicans now lead by 1.7 points; curiously, that is exactly the margin by which Republicans lead if I assume 1.5% vote third-party and undecided voters split 2-1 Republican. By one model, this translates to a gain by Republicans of 16 seats in the United States House of Representatives in 2022, giving them a 229-206 majority. Some VERY back of the envelope math using R+1.7 also suggests a net gain of one seat for Republicans in the United States Senate in 2022, giving them control there as well.

I will let others decide how Democrats should counteract these possible trends – though I will observe that political trends follow sine curves, not arrows – and instead conclude that public polling has profound consequences, both in the broad picture it paints – blowout lowering turnout vs. nail-biter increasing turnout – and in its accuracy – it cannot continue to consistently undercount Republican voters or even hardcore analysts like me will fundamentally question its utility.

Until next time…please wear a mask as necessary to protect yourself and others – and if you have not already done so, get vaccinated against COVID-19! And if you like what you read on this website, please consider making a donation. Thank you.


[1] Republican Tom Kean won in 1981 and 1985. Democrat Jim Florio won in 1989, then lost very narrowly to Christine Todd Whitman in 1993. Whitman was reelected in 1997 by an equally-narrow margin over Democrat Jim McGreevey. McGreevey then won easily in 2001, only to resign because of a sex scandal in August 2004. Democrat John Corzine won easily in 2005 (suggesting McGreevey likely would have been reelected), then lost to Republican Chris Christie in 2009. Christie cruised to reelection in 2013, only to see Murphy win in 2017. For gubernatorial election data going back to 1993, see Dave Leip’s indispensable Atlas of U.S. Presidential Elections.

[2] The second and final Rutgers University poll had Murphy up 8 points (10/21-10/27).

[3] Data from a personal dataset of polls from that election.

One thought on “2021 Election Post-Mortem: How much does polling drive turnout?

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s