2020 Elections Post-Mortem

On November 3, 2020, the United States ended a weeks-long electoral process. At stake was the presidency, control of the United States Senate (“Senate”) and House of Representatives (“House”), 11 governor’s mansions, and thousands of state and local offices. That day, I published “cheat sheets” to guide election viewers through state-level presidential returns, 35 Senate elections and the gubernatorial elections.

[Ed. note: This post, my 200th, is the longest I have written to date. It is fitting that a blog which found its data-driven footing in the wake of the 2016 elections would have its 200th entry address the aftermath of the 2020 elections, beyond mere repetition of the number “20.”]

As I write this on midnight EST on November 17, 2020, precisely two weeks after the elections concluded, these are the top-line results:

  • Only one governor’s mansion changed partisan hands: Republican Greg Gianforte won back the statehouse in Montana for the first time in 16 years. As of January 2021, Republicans will hold 27 governor’s mansions, and Democrats will hold 23.
  • Democrats basically held serve in state legislative races. For more details, please see here.

On balance, the 2020 elections affirmed the status quo: a nation roughly evenly split between Democrats and Republicans, though it remains possible the former could control, however narrowly, the White House, Senate and House for the first time since 2010.

Presidential election

Biden-Harris are closing on 79.0 million votes (50.9%), shattering the previous record of 69.5 million votes won by Democrat Barack Obama and Biden in 2008. Trump-Pence have just under 73.3 million votes (47.3%), ranking them second in history. Biden has now appeared on three of the four presidential tickets to receive the most votes, with Obama-Biden earning 65.9 million votes in 2012, edging out Clinton-Kaine in 2016 by about 65,000 votes. Third party candidates are receiving more than 2.8 million votes (1.8%), significantly lower than the 8.3 million votes (6.0%) such candidates received in 2016. Approximately 155.1 million votes have already been counted, with an estimated 4.1 million votes—mostly in California and New York—left to be counted. This ~159.2 million vote projection, or about 2/3 of all Americans eligible to vote, also shatters the previous record of 137.1 million votes set in 2016.

Biden-Harris’ 3.6 percentage point (“point”) margin is a 1.5-point increase from 2016, and 0.3-point decrease from 2012, making it the third consecutive presidential election in which the Democratic ticket won the national popular vote by between two and four points; adding 22 million voters did not fundamentally alter the partisan electoral divide. Based on my Electoral College model, a Biden-Harris win of 3.6 points equates to 296 EV, nearly the 306 EV they received; for a Republican ticket, this equates to 327 EV.

How did Biden-Harris win the Electoral College?

Table 1: 2020 and 2016 Presidential Election Results by State, Ranked from Highest to Lowest Biden-Harris Margin

StateEVWinnerClinton-Kaine MarginBiden-Harris MarginDelta
DC3Biden86.886.6-0.2
Vermont3Biden26.435.49.0
Massachusetts11Biden27.233.05.8
Maryland10Biden30.032.52.5
California55Biden26.429.63.2
Hawaii4Biden32.229.5-2.7
Rhode Island4Biden15.520.85.2
Connecticut7Biden13.620.16.7
Washington12Biden15.719.33.6
Delaware3Biden11.319.07.7
Illinois20Biden14.016.62.6
Oregon7Biden11.016.25.2
New Jersey14Biden16.915.5-1.4
New York29Biden22.513.7-8.8
Colorado9Biden4.913.58.6
New Mexico5Biden8.210.82.6
Virginia13Biden3.010.16.9
Maine4Biden (3)5.38.73.4
New Hampshire4Biden0.47.47.0
Minnesota10Biden1.57.15.6
Michigan16Biden-0.22.62.8
Nevada6Biden2.42.40.0
Pennsylvania20Biden-0.71.01.7
Wisconsin10Biden-0.80.61.4
Georgia16Biden-5.10.35.4
Arizona11Biden-3.50.33.8
North Carolina15Trump-3.7-1.42.3
Florida29Trump-1.2-3.4-2.2
Texas38Trump-9.0-5.73.3
Ohio18Trump-8.1-8.2-0.1
Iowa6Trump-9.4-8.21.2
Alaska3Trump-14.3-10.13.3
South Carolina9Trump-20.4-11.78.7
Kansas6Trump-18.5-15.13.4
Missouri10Trump-19.0-15.63.4
Indiana11Trump-20.2-16.14.1
Montana3Trump-14.7-16.4-1.7
Mississippi6Trump-25.1-17.87.3
Louisiana8Trump-19.6-18.61.0
Nebraska5Trump (4)-17.8-19.2-1.4
Utah6Trump-17.9-20.2-2.3
Tennessee11Trump-31.8-23.38.5
Alabama9Trump-26.0-25.60.4
Kentucky8Trump-27.7-26.01.7
South Dakota3Trump-29.8-26.23.6
Arkansas6Trump-29.8-27.62.2
Idaho4Trump-26.9-30.8-3.9
Oklahoma7Trump-36.4-33.13.3
North Dakota3Trump-35.7-33.42.4
West Virginia5Trump-41.7-39.02.7
Wyoming3Trump-46.3-43.42.9
Average  Trump+3.6Trump+0.8D+2.8

As Table 1 reveals, Biden-Harris won 25 states and the District of Columbia (“DC”) by an average of 17.4 points, while Trump-Pence won 25 states by an average of 19.8 points; medians are 14.6—reflecting the 86.8-point margin in DC—and 18.6, respectively. Biden-Harris won seven states and DC totaling 97 EV by 20 or more points, while Trump-Pence won 11 states totaling 65 EV by that margin.

Biden-Harris won 19 states, DC and the 2nd Congressional district in Nebraska by at least 6.0 points, for a total of 228 EV. Add Nevada (6) and Michigan (16), which the Democratic ticket won by ~2.5 points, below their national margin, and the total increases to 250 EV.

At around 10:30 am EST on Saturday, November 7, the major news networks declared Biden-Harris the projected winner in Pennsylvania—and its 20 EV put Biden-Harris over the total of 270 needed to win the presidency. It also makes Pennsylvania—the state in which I was born—the “tipping point” state, as it puts Biden-Harris over 270 EV when states are ranked from most to least Democratic. But the margin stands at just 1.0 points, or just 68,903 votes; Biden-Harris also won Wisconsin (0.6 points), Arizona and Georgia (0.3 points each) by similarly small margins. The Democratic ticket has a total winning margin of 104,025 votes in these four states.

In the 25 states, plus DC, won by the Democratic ticket, the average increase in margin from 2016 was 3.4 points, while in states won by the Republican ticket the average increase was 2.1 points; overall, the average margin shift was 2.8 points. In the five states which switched from Republican to Democratic, the average increase was 3.0 points, led by a 3.8-point increase in Arizona and a 5.4-point increase in Georgia. While Biden-Harris lost North Carolina by 1.4 points and Texas by 5.7 points, they improved the margin by 2.3 and 3.3 points, respectively.

However, while Biden-Harris improved on the 2016 margins by an average 3.7 points in these four southeastern/southwestern states—states I suggested were fertile ground for Democrats—they basically held serve in Iowa (D+1.2) and Ohio (no change), while falling further behind in Florida (D-2.2); I will not speculate what role undelivered ballots in Miami-Dade County played in the latter state. This should not be surprising, as these were perhaps the most disappointing states for Democrats during the otherwise “blue wave” 2018 midterm elections.

In 2016, Trump-Pence won 306 EV by winning six states Obama-Biden won in 2012: the aforementioned Florida, Iowa and Ohio, plus Michigan, Pennsylvania and Wisconsin. The latter were decided by a combined 77,736 votes; Clinton-Kaine also lost Georgia by 211,141 votes and Arizona by 91,234 votes. In 2020, as Table 2 shows, Biden-Harris won the former three states—more than enough to give them an Electoral College victory—by a combined 233,945 votes: a shift of 311,681 votes, or just 0.2% of all votes cast. But the Democratic ticket also increased their margin in Arizona by 101,691 votes and in Georgia by a remarkable 226,296 votes.

Table 2: Changes in Margin from 2016 to 2020 in Five Key States

State2016 Dem Margin2020 Dem MarginIncrease, 2016-20
Michigan-10,704+144,532+155,236
Pennsylvania-44,284+68,903+113,187
Wisconsin-22,748+20,510+43,258
Arizona-91,234+10,457+101,691
Georgia-212,141+14,155+226,296
TOTAL-381,111+258,557+639,668

Overall, across these five states, the margin swung toward the Democratic ticket by about 640,000 votes, which is still less than 1% of all votes cast. But we can get even more granular than that. Early in 2017, I observed that in the three states that swung the 2016 election to Trump-Pence, the Clinton-Kaine ticket did about as well in the Democratic core counties—the urban centers of Detroit, Milwaukee/Madison and Philadelphia/Pittsburgh—as Obama-Biden had in 2012. What changed was a massive increase in Republican turnout in the other, more rural counties of those states. I ultimately concluded this resulted from a split between white voters with a college degree (more Democratic suburban/urban) and without a college degree (more Republican rural).

Table 3: Changes in Margin from 2016 to 2020 in Pennsylvania Counties

County2016 D Margin2020 D MarginIncrease, 2016-20
Phila Suburbs   
Bucks+2,699+17,415+14,716
Chester+25,568+53,598+28,030
Delaware+66,735+87,066+20,331
Montgomery+93,351+133,343+39,992
TOTAL+188,353+291,422+103,069
    
Major Urban   
Philadelphia+475,277+457,649-17,628
Allegheny+108,137+146,706+38,569
TOTAL+583,414+604,355+20,941
    
All Other Counties-816,051-826,874-10,283
TOTAL-44,284+68,903+113,187

Table 3 shows just how this split played out in 2020, using Pennsylvania as an example. Compared to 2016, the margins for the Democratic ticket increased only at 21,000 votes in the heavily urban Democratic counties of Allegheny (Pittsburgh) and Philadelphia. And the 61 counties outside these two counties, excepting the four-county ring around Philadelphia, also held serve for the Republican ticket; Erie and Northampton Counties switched, barely, from Trump-Pence to Biden-Harris. In fact, the two parties may have reach voted saturation in these two areas. But those four suburban Philadelphia counties, swung even further toward the Democratic ticket, from a margin of 188,353 votes to nearly 291,422 votes, for a total increase of 103,069 votes, nearly the entire swing from 2016 to 2020.

What about the polling?

With most of the vote counted, Biden-Harris lead Trump-Pence nationally by 3.6 points, which is 4.6 points lower than my final weighted-adjusted polling average (“WAPA”) of 8.2 points.

For my final post tracking national and state polling of the 2020 presidential election, I estimated the probability Biden-Harris would win a given state. In 24 states/DC totaling 279 EV, the probability was at least 94.7%; Biden-Harris won all of them. In 20 states totaling 126 EV, the probability was 1.3% or less; Trump-Pence won all them. The remaining seven states were:

  • Florida (80.1%), which Biden-Harris lost
  • Arizona (77.5%), which Biden-Harris won
  • North Carolina (69.0%), which Biden-Harris lost
  • Georgia (56.4%), which Biden-Harris won
  • Ohio (39.1%), Iowa (37.0%) and Texas (28.4%), each of which Trump-Pence won

Florida and North Carolina were the only “misses,” though it should be noted Trump-Pence still had a non-trivial 19.9% and 31.0% chance, respectively, to win those states. Further, my final back-of-the-envelope EV estimate was 348.5 for Biden-Harris—subtracting the 44 combined EV of Florida and North Carolina essentially gets you to 306. The latter value is also very close to the 297.5 EV I estimated Biden-Harris would receive if all polls overestimated Democratic strength by 3.0 points.

Along those lines, my 2020 election cheat sheets included a projected Democratic-minus-Republican margin (“JBWM”), which adjusts final WAPA for undecided votes, along with recent polling errors in selected states. Compared to the final FiveThirtyEight.com margins/polling averages (“538”), JBWM margins were about 1.2 points more Republican.

Even so, as Table 4 shows, the JBWM margins were, on average, 3.4 points more Democratic than the final margins, and the 538 margins were 4.6 points more Democratic. When the direction of the difference is ignored, meanwhile, the differences between the two method vanish: an average absolute difference of 4.5 from JBWM margins compared to 4.8 for 538.

However, this overall difference masks a stark partisan difference: the mean JBWM difference was only 1.1 points more Democratic in states/DC won by Biden-Harris, while it was 5.9 points more Democratic in states won by Trump-Pence. The correlation between the Biden-Harris margin and the JBWM difference is 0.73, meaning the more Republican the state, the better Trump-Pence did relative to the final polling. In short, pollsters continue to undercount “Trump Republicans” in the most Republican states.

Table 4: 2020 Presidential Election Results by State, Ranked by Difference from JBWM Democratic-Republican Margin “Projection”

StateEVWinnerJBWM ProjectionBiden-Harris MarginDelta
West Virginia5Trump-20.4-39.0-18.6
New York29Biden28.313.7-14.6
Wyoming3Trump-32.1-43.4-11.3
South Dakota3Trump-15.6-26.2-10.6
North Dakota3Trump-23.2-33.3-10.1
Montana3Trump-7.1-16.4-9.3
Kentucky8Trump-17.2-26.0-8.8
Oklahoma7Trump-24.9-33.1-8.2
Texas38Trump1.6-5.7-7.3
Utah6Trump-12.9-20.2-7.3
Alabama9Trump-18.6-25.6-7.0
Indiana11Trump-9.6-16.1-6.5
Tennessee11Trump-16.9-23.3-6.4
Nevada6Biden8.62.4-6.2
Missouri10Trump-9.6-15.6-6.0
Kansas6Trump-10.2-15.1-4.9
Idaho4Trump-26.0-30.8-4.8
New Jersey14Biden19.515.5-4.0
Maine4Biden (3)12.58.7-3.8
Mississippi6Trump-14.1-17.8-3.7
Florida29Trump0.2-3.4-3.6
Alaska3Trump-6.7-10.1-3.4
Iowa6Trump-5.0-8.2-3.2
Connecticut7Biden23.020.1-2.9
Louisiana8Trump-15.8-18.6-2.8
South Carolina9Trump-8.9-11.7-2.8
Wisconsin10Biden3.20.6-2.6
Arizona11Biden2.90.3-2.6
Washington12Biden21.719.3-2.4
Hawaii4Biden31.629.5-2.1
Ohio18Trump-6.2-8.2-2.0
Michigan16Biden4.42.6-1.8
New Hampshire4Biden8.97.4-1.5
Nebraska5Biden (4)-17.8-19.2-1.4
Massachusetts11Biden34.333.0-1.3
Oregon7Biden17.516.2-1.3
New Mexico5Biden12.010.8-1.2
Pennsylvania20Biden2.21.0-1.2
Delaware3Biden20.219.0-1.2
Virginia13Biden11.210.1-1.1
Minnesota10Biden7.97.1-0.8
Georgia16Biden0.60.3-0.3
North Carolina15Trump-1.1-1.4-0.3
Illinois20Biden16.216.60.4
California55Biden29.129.60.5
Arkansas6Trump-29.1-27.61.5
Maryland10Biden30.231.91.7
Rhode Island4Biden19.020.71.7
Colorado9Biden11.413.52.1
Vermont3Biden28.835.46.6
DC3Biden74.986.611.7
Average  Biden+2.6Trump+0.8D-3.4

To again get more granular, Table 5 lists the pollsters who assessed the national popular vote at least five times since January 1, 2019, sorted by distance from the actual national margin of 3.6%. Margins are weighted for time, but not adjusted for partisan “bias.”

Table 5: Top 2020 Presidential Election Pollsters, Final WAPA National Margin

Pollster538 RatingFinal MarginDelta
OpiniumC+14.1-10.5
NORC (AllAdults only)C+11.3-7.7
CNN/SSRSB/C11.1-7.5
QriouslyC+10.5-6.9
USC DornsifeB/C10.4-6.8
Quinnipiac UniversityB+10.4-6.8
NBC News/Wall Street JournalA-10.1-6.5
Global Strategy Group/GBAO (Navigator Res)C+9.9-6.3
Data for ProgressB-9.8-6.2
Redfield & Wilton StrategiesC+9.6-6.0
ABC News/Washington PostA+9.2-5.6
Marist CollegeA+9.1-5.5
Echelon InsightsC+8.8-5.2
SurveyUSAA8.8-5.2
IpsosB-8.5-4.9
LégerC+8.4-4.8
Change ResearchC-8.3-4.7
Fox NewsA-8.3-4.7
YouGovB8.2-4.6
Research Co.B-7.8-4.2
PureSpectrumC+7.6-4.0
Morning ConsultB/C7.6-4.0
Monmouth UniversityA+7.4-3.8
Firehouse Strategies/OptimusB/C7.4-3.8
RMG ResearchB/C7.1-3.5
Harris XC6.5-2.9
Suffolk UniversityA6.2-2.6
IBD/TIPPA/B5.5-1.9
Emerson CollegeA-3.8-0.2
Zogby*C+3.60.0
Rasmussen Reports/Pulse Opinion ResearchC+3.20.4
CiviqsB/C3.10.5
AverageB/B-8.2-4.5

           * John Zogby Strategies/EMI Solutions, Zogby Analytics, Zogby Interactive/JV Analytics

These 32 pollsters accounted for 556 (80.6%) of the 690 polls conducted. On average, they estimated Biden-Harris would win the national popular vote by 8.2 points, identical to my final WAPA; the average miss was 4.5 points in favor of Biden-Harris. There was only minimal difference by pollster quality: the 11 pollsters with a rating of B or better missed by an average of 4.2 points, while the 21 pollsters with a rating of B- or lower missed by an average of 4.7 points. That said, three of the four pollsters who came closest to the final national margin—Zogby, Rasmussen and Civiqs—had ratings of B/C or C+; the fourth was Emerson College, rated A-. At the other end of the spectrum are seven pollsters who anticipated a double-digit national popular vote win for Biden-Harris: low-rated Opinium, NORC (who polled adults, not registered/likely voters), CNN/SSRS, Qriously and USC Dornsife; and high-rated Quinnipiac University and NBC News/Wall Street Journal.

Overall, though, the polling captured the broad contours of the 2020 presidential election—if not the precise margins—fairly well, with JBWM and actual Democratic margins correlated a near-perfect 0.99; the order of states from most to least Democratic was accurately predicted. It forecast a solid, if not spectacular win by Biden-Harris in the national popular vote, a restoration of the upper Midwestern “blue wall,” and continued Democratic gains in southeastern/southwestern states such as Arizona, Georgia, North Carolina and Texas, even as Florida, Iowa and Ohio become more Republican.

One final note: it is exceedingly difficult to beat an elected incumbent president. Since 1952, it had happened only twice (1980, 1992) in eight chances prior to 2020[1]; Biden-Harris beat those 1:3 odds convincingly.

Senate elections

Democrats entered 2020 needing to flip a net four seats—or three seats plus the White House—to regain the majority for the first time since 2014. As Vice-President-elect, Kamala Harris breaks a 50-50 tie.

Table 6: 2020 Senate Election Results by State, Ranked from Highest to Lowest Democratic Margin, Compared to Pre-Election “Fundamentals”

StateWinnerFundamentalsFinal Dem MarginDelta
Rhode IslandReed24.433.08.6
MassachusettsMarkey28.532.94.4
DelawareCoons18.921.52.6
OregonMerkley15.117.62.5
IllinoisDurbin21.116.9-4.2
New JerseyBooker18.415.8-2.6
New HampshireShaheen6.515.79.2
VirginiaWarner7.912.04.1
ColoradoHickenlooper1.89.37.5
New MexicoLujan8.56.1-2.4
MinnesotaSmith5.75.2-0.5
ArizonaKelly-8.32.410.7
MichiganPeters8.61.5-7.1
Georgia Special???-8.0-1.07.0
Georgia???-10.0-1.78.3
North CarolinaTillis-6.4-1.74.7
IowaErnst-5.1-6.6-1.5
MaineCollis5.5-8.9-14.4
TexasCornyn-15.7-9.85.9
MontanaDaines-19.0-10.09.0
South CarolinaGraham-16.1-10.35.8
MississippiHyde-Smith-18.1-11.26.9
KansasMarshall-21.4-11.99.5
AlaskaSullivan-19.6-12.96.7
KentuckyMcConnell-29.1-19.59.6
AlabamaTuberville-24.2-20.63.6
LouisianaCassidy-22.6-25.9-3.3
TennesseeHagerty-23.8-27.1-3.3
IdahoRisch-34.6-29.35.3
OklahomaInhofe-38.5-30.28.3
South DakotaRounds-26.2-31.5-5.3
ArkansasCotton-28.6-33.3-4.7
NebraskaSasse-26.2-41.3-15.1
West VirginiaCapito-35.9-43.3-7.4
WyomingLummis-43.7-46.1-2.4
AverageD+1 to 3GOP+8.9GOP+7.0D+1.9

Table 1 summarizes these elections; for the Georgia special election and Louisiana, margins are for all Democrats and all Republicans. Democrats John Hickenlooper and Mark Kelly defeated Republican incumbents in Colorado (Cory Gardner) and Arizona (Martha McSally), respectively, while Republican Tommy Tuberville defeated Democratic incumbent Doug Jones in Alabama. This leaves Democrats two seats shy of 50-50, pending the January 5 runoff elections in Georgia. Incumbent Republican David Perdue edged Ossoff on November 3 by 1.7 points, but fell 0.3 points short of the 50.0% needed to win outright. In the special election necessitated by the retirement of Republican Johnny Isaakson in December 2019, Warnock (32.9%) led incumbent Republican Kelly Loeffler (25.9%) by 7.0 points in the all-candidate “jungle primary;” overall, Republican candidates earned 49.4% of the vote and Democratic candidates earned 48.4%, with 2.2% split between a handful of third-party candidates.

When I took a “wicked early” look at these elections, I assessed the Democrat’s chance in each election using their “fundamentals,” or the sum of the state’s partisan lean (calculated using my 3W-RDM), the Democratic margin on the generic ballot and incumbency advantage.[2] For Table 6, the generic ballot is the difference in the percentages of the total vote for all Democratic House candidates and for all Republican House candidates; Democrats are ahead by 2.0 points.

On average, Democrats overperformed “expected” margins by 1.9 points. In the 13 elections won by Democrats, the overperformance was 2.5 points, while in the 20 elections won by Republicans, the overperformance was just 0.9 points; Democrats overperformed in the two Georgia Senate races by 7.0 and 8.3 points, confirming how rapidly it is moving toward swing-state status. The biggest Democratic overperformance—fully 10.7 points—was in Arizona, which in 2021 will have two Democratic Senators (both of whom beat McSally) for the first time since 1953. Other Senate elections in which the Democratic candidate overperformed by at least 9.0 points were New Hampshire, and three states where Democrats fell short in their attempt to win back a Republican-held seat: Montana, Kansas and Kentucky.

On the flip side, setting aside a 15.1-point underperfomance in Nebraska, the biggest Republican overperformance was in Maine, where incumbent Susan Collins, first elected in 1996, “should” have lost by 5.5 points. Instead, she won by 8.9 points; this is a 28-point decline from 2014, when Collins won by 37 points. Pending the results of the Georgia runoff elections, Maine is the only state in 2020 to have a Democratic presidential victory and a Republican Senate victory, with a gap of 17.6 points. It will be interesting to see whether Collins adjusts her voting in the next Senate. Other large Democratic underperformances, finally, took place in Michigan, where first-term Democratic Senator Gary Peters beat Republican John James by only 1.5 points and in West Virginia, which grows more Republican every year.

On the whole, though, expected and actual margins aligned nearly perfectly, with a 0.94 correlation.

What about the polling?

As with the presidential election, the final polling averages/projected margins were far less accurate, as Table 7 shows; I did not calculate a projected final margin for the Louisiana Senate election.

Table 7: 2020 Senate Election Results by State, Ranked by Difference from JBWM Democratic-Republican Margin “Projection”

StateWinnerJBWM ProjectionDemocratic MarginDelta
West VirginiaCapito-20.6-43.3-22.7
WyomingLummis-30.2-46.1-15.9
MaineCollins3.3-8.9-12.2
South DakotaRounds-19.9-31.5-11.6
NebraskaSasse-30.8-41.3-10.5
KentuckyMcConnell-9.7-19.5-9.8
OklahomaInhofe-20.5-30.2-9.7
AlaskaSullivan-3.7-12.9-9.2
AlabamaTuberville-11.5-20.6-9.1
New JerseyBooker24.615.8-8.8
MontanaDaines-1.3-10.0-8.7
DelawareCoons29.621.5-8.1
TexasCornyn-2.3-9.8-7.5
IllinoisDurbin23.716.9-6.8
KansasMarshall-5.4-11.9-6.5
South CarolinaGraham-4.7-10.3-5.6
TennesseeHagerty-21.9-27.1-5.2
MississippiHyde-Smith-6.4-11.2-4.8
ArizonaKelly6.62.4-4.2
New MexicoLujan10.06.1-3.9
Georgia???1.7-1.7-3.4
MichiganPeters4.71.5-3.2
IdahoRisch-26.1-29.3-3.2
MinnesotaSmith8.45.2-3.2
VirginiaWarner15.112.0-3.1
IowaErnst-3.6-6.6-3.0
North CarolinaTillis1.1-1.7-2.8
OregonMerkley20.017.6-2.4
ArkansasCotton-33.0-33.3-0.3
ColoradoHickenlooper9.39.30.0
MassachusettsMarkey31.432.91.5
New HampshireShaheen14.415.71.3
Georgia Special???-3.9-1.02.9
Rhode IslandReed29.633.03.4
AverageDem+1 to 3GOP+0.6GOP+6.4D-5.8

The polling may have been within historic parameters for the presidential election, but it was far worse in the Senate elections, with the JBWM margins overestimating Democratic margins by an average of 5.8 points, almost exactly the 6.0 points by which 538 margins erred on average; ignoring direction, the average misses are 6.3 and 7.0 points, respectively. That said, the correlation between the actual and projected Democratic margins was 0.97, meaning the polling correctly forecast the order of Senate elections from most to least Democratic.

These overall averages again mask substantial partisan differences. In the 13 states where the Democratic nominee won, the average miss was a historically-reasonable -2.9 points, but in the 19 states (excluding Louisiana) where the Republican nominee won, the average miss was an astounding -8.3 points. Put another way, in the 15 states Trump-Pence won by at least 10 points which also held a Senate election, the average Senate miss was -8.9 points, while it was -3.3 points in all other states. Somewhat reassuringly, in the five states whose presidential margin was within five points also holding a Senate election (Arizona, Georgia, Michigan, North Carolina), the miss was only -2.2 points. Overall, the correlation between the Biden-Harris margin and the JBWM margin error was 0.57, confirming the idea pollsters systematically undercounted Republican support in the most Republican states.

My back-of-the-envelope estimate was a net gain of five Democratic seats in the Senate, with at least a 77% chance Democrats would regain control; these values dropped to 30% and either two or three seats with the assumption all polls systematically overestimated Democratic strength by three points. Democrats will ultimately net between one and three seats, corresponding more with the latter assumption. I estimate the probability Democrats win both Georgia Senate runoff elections—and thus the Senate—is between 25 and 50%, depending on the degree of ticket-splitting.

From a purely mathematical perspective, the largest Democratic underperformances occurred in the Senate elections in West Virginia, Wyoming, South Dakota and Nebraska: four extremely Republican states. But from a strategic perspective, the most disappointing elections were in Maine (-12.2) and North Carolina (-2.8), where incumbent Republican Thom Tillis narrowly held off a challenge from Democrat Cal Cunningham, who may have been hurt by a sexting scandal; given the narrowness of his victory (1.7 points) and the increasingly swing status of North Carolina, Tillis’ voting patterns also merit watching. These were the two states besides Arizona (98.1%) and Colorado (99.5%) in which I estimated the Democratic nominee had at least an 85% chance to defeat a Republican incumbent; I also thought Democrat Theresa Greenfield was roughly even money to defeat incumbent Republican Joni Ernst, despite projecting a final margin of 3.6 points for Ernst; the latter won by 6.6 points.

There were four additional Senate elections—in Alaska, Kansas (open seat), Montana and South Carolina—where I estimated the probability of a Democratic flip was between 11.7 and 26.4%. In a sign of how good these elections were for Republicans, their nominees won all four elections by an average of 11.3 points, a mean 7.5 points more Republican than projected. In fairness, these states tilted an average 19.2 points more Republican than the nation as a whole coming into the 2020 elections. A similar story can be told in Texas, which tilted 15.3 points more Republican, but where Democrat M.J. Hegar “only” lost by 9.8 points to incumbent Republican John Cornyn, beating expectations by 0.6 points.

Put simply, assuming a loss in Alabama, Democratic hopes of winning back control of the Senate relied on flipping two Senate seats in Democratic states, then winning at least two more seats in states ranging from somewhat Republican—Iowa, North Carolina, Arizona, Georgia—to extremely Republican—Alaska, Kansas, Montana, South Carolina and Texas—all while Trump sought reelection. To date, Democrats have only flipped seats in Colorado (D+2.2) and Arizona (D-9.7) while winning back the Vice-Presidency, losing tough elections in Iowa, Maine and North Carolina, while never really being in contention anywhere else. Senate control now rests on Democrats winning two Senate runoff elections in a nominally Republican state (D-9.6), but one where Biden-Harris won, improving on Clinton-Kaine’s by 5.4 points.

Gubernatorial elections

Unlike those for the White House and Senate, there was very little drama in these elections. Two Democratic incumbents—John Carney of Delaware and Jay Inslee of Washington—were expected to win easily; they won by margins of 20.9 and 13.6 points, respectively. Six Republican incumbents—Eric Holcomb of Indiana, Mike Parson of Missouri, Chris Sununu of New Hampshire, Doug Burgum of North Dakota, Phil Scott of Vermont and Jim Justice of West Virginia—as well as Republican Spencer Cox of Utah were expected to win easily, though I projected Parson to win by “only” 8.0 points (he won by 16.6 points). They won their elections by an average margin of 31.6 points!

The only possible drama was in Montana, where Republican Gianforte and Democrat Mike Cooney vied to win the governor’s mansion being vacated by Democrat Bullock, and North Carolina, where Democratic Governor Roy Cooper—who won extremely narrowly in 2016—faced Republican Dan Forest. Gianforte defeated Cooney by 12.4 points, easily exceeding a projected 4.5 points, while Cooper won by 4.5 points, not the projected 10.4 points. Still, my global projection was correct: a net gain of one governor’s mansion by the Republicans, giving them a 27-23 majority; this an overall net gain of seven governor’s mansions by the Democrats since 2016.

In these elections, Republicans strongly overperformed fundamentals (7.1 points) and JBWM projections (7.6 points). However, expected values were strongly skewed by Scott’s 41.1-point victory in extremely-Democratic Vermont (D+27.7) and Sununu’s 31.8-point victory in swing New Hampshire (D+0.1); exclude those two margins and DEMOCRATS overperformed expectations by 1.0 points—with Democrat Ben Salango exceeding what were admittedly very low expectations by 8.5 points. Meanwhile, in the four states with governor’s races won by Biden-Harris, Democratic gubernatorial nominees finished an average 8.9 points lower than projected, while in the seven states won by Trump-Pence, they finished an average 6.8 points worse than expected. Once again, the extreme disparity in presidential/Senate and gubernatorial voting in New Hampshire and Vermont—two of three states in solidly-Democratic New England, along with Massachusetts (Charlie Baker), to have very popular Republican governors. In fact, gubernatorial elections are among the only ones in which ticket-splitting is still relatively common: Biden-Harris won six states with a Republican governor,[3] while Trump-Pence won five states with a Democratic governor.[4]

House elections

A wide range of forecasters expected Democrats to net between five and 10 House seats[5]. I was highly dubious of this, to be honest, given the likelihood the margin for Democrats in the total national House vote would decline from the 8.6-point margin they earned in 2018; it would also be higher than the 1.1 points by which they lost in 2016, when they still managed to net six seats. However, because I was not closely tracking House races, I said nothing about my doubts.

According to the Cook House vote tracker, Democrats had earned more than 75.1 million House votes (50.1%), Republicans had earned just under 72.1 million votes (48.0%), with the nearly 2.2 million votes (1.8%) going to third-party candidates. A total of 150.0 million votes have been counted, 5.1 million less than those cast in the presidential election. The 2.0-point margin by which Democrats are winning the House vote—just under 3.1 million votes—is also lower than the 3.6 points, and 5.6 million votes by which Biden-Harris currently lead Trump-Pence. It is also much lower than the 9.7-million Democratic vote margin in 2018, albeit with 36.3 million more votes cast in 2020, reinforcing the conclusion a few million Republican-leaning voters “balanced” a vote for Biden-Harris with Republican votes elsewhere…or simply chose not to vote in down-ballot elections.

In the races that have already called, Republicans have gained 11 seats held by Democrats (two each in California and Florida, one each in Iowa, Minnesota, New Mexico, New York, Oklahoma, South Carolina and Utah), while Democrats have gained three seats held by Republicans (two in North Carolina, one in Georgia). This gives Democrats 221 seats, three more than needed for the majority, and Republicans 208 seats. Of the six seats yet to be called, Democrats currently hold four, with freshman Democrat Tom Malinowski leading by ~5,000 votes in New Jersey’s 7th Congressional District (“CD”). Giving that seat to the Democrats—and giving Republicans their open seat in New York’s 2nd CD—increases the totals to 222 Democrats and 209 Republicans.

That leaves four seats truly in doubt:

  • California’s 21st CD, where incumbent Democrat T.J. Cox trails Republican David Valadao, in a 2018 rematch, by 2,065 votes.
  • California’s 25th CD, where Democrat Christy Smith is within 104 votes of unseating Republican Mike Garcia, who won a special election in May 2020 after first-term Democrat Katie Hill resigned.
  • Iowa’s 2nd CD, where Democrat Dave Loebsack did not seek reelection; Republican Mariannette Miller-Meeks leads Democrat Rita Hart by only 47 votes!
  • New York’s 22nd CD, where Republican Claudia Tenney’s lead over incumbent Democrat Anthony Brindisi continues to shrink as New York votes are slowly counted.

Democrats will thus lose a net 8-12 seats compared to the 234-201 margin they had after the 2018 elections. This is a bad result for the Democrats, right?

Well, no…it suggests that polling-based expectations were flawed, because the fundamentals always pointed toward a net loss of House seats for the Democrats. Moreover, the comparison should be to 2016, because that is the last election in which Trump appeared on the ballot.

Following the 2016 elections, Republicans had a 241-194 House majority. Democrats were convinced, wrongly I thought, that gerrymandering by Republican legislators and governors would keep them in the minority for the foreseeable future. Looking ahead to the 2018 midterm elections, knowing Democrats needed to net 24 seats to regain the majority, I looked at all House elections from 1968 to 2016, and I noticed that what “predicted” net change in seats from one election to the next was not the national margin in a given election, but the change in that margin from the previous election. Figure 1 helps to illustrate this.

Figure 1:

In 2018, Democrats net a surprisingly-high 41 House seats, 17 more than they needed, most by narrow margins. It is then reasonable to expect that even a small decline in the Democratic share of the total national House vote would allow Republicans to “claw back” some of these seats Democrats currently lead the total national House vote by 2.0 points, fully 6.6-point decrease f 2018. Entering this value into the OLS regression shown in Figure 1 yields an estimated Democratic loss of 22.4 seats.

In other words, while Democrats expected to gain seats—based on what we now know was polling that underestimated Republican margins by 3-7 points—they should actually have been bracing themselves for a possible loss of the House itself. Instead, they “only” lost between eight and 12 seats, meaning they did far better than history would have suggested. Moreover, Democrats have net between 29 and 33 seats since 2016, earning control of the House in back-to-back elections for the first time since 2006-2008, something that seemed nearly impossible early in 2017.

Summary

Both Democrats and Republicans can find 2020 election results to celebrate.

Democrats won back the White House after just four years (beating 1:3 odds to defeat an incumbent), rebuilding their upper-Midwestern blue wall while expanding into the southeast and southwest; no Democratic presidential nominee has won both Arizona and Georgia since 1948. They also maintained control of the House of Representatives and made gains in the Senate; with two more wins in Georgia in January 2021, they regain control of the Senate as well. Democrats have not controlled both the White House and House since 2010.

Republicans, even as they lost the White House, gained as many as 12 seats in the House and staved off losing control of the Senate until January 2021 at the earliest. They net one governor’s mansion, giving them a 27-23 majority, and held their own in state legislative elections. Once again, Trump’s name on the ballot encouraged many more exurban and rural voters to vote than expected, ironically helping all Republicans but himself and his running mate.

Fans of bipartisan “balance” can also celebrate 11 states seeing different parties win their state’s electoral votes and serving as governor. Moreover, a record-smashing 155.1 million—and counting—Americans cast a ballot for president, which equates to two in three of all adults eligible to vote.

Finally, the polls erred substantially in favor of Republicans, with a miss of around 3.5 points compared to my final projections and 4.7 points relative to those from 538. Republicans fared even better in Senate and gubernatorial elections, beating final projections by around six points in the former and nearly eight points in the latter. These values mask a partisan split, with polls far more accurate for Democratic candidates than Republican ones. In the end, though, polls were far less accurate—in this Trump-led cycle at least—than simply considering a state’s recent partisan lean, the national partisan environment and incumbency. These fundamentals remain extremely predictive, at least relatively.

Until next time…please stay safe and healthy…


[1] 1956, 1972, 1980, 1984, 1992, 1996, 2004, 2012

[2] Democratic full-term incumbents=4.4, Democratic partial-term incumbents=2.2, non-incumbent=0, Republican partial-term incumbents=–0.4, -0.6, -1.6; Republican full-term incumbents=-2.4

[3] Arizona, Georgia, Maryland, Massachusetts, New Hampshire, Vermont

[4] Kansas, Kentucky, Louisiana, Montana, North Carolina

[5] The Cook Political Report hedged a bit, labeling 229 seats at least Lean Democrat, 179 seats at least Lean Republican, and 27 seats Toss-up. Of the Toss-ups, nine are held by Democrats, 17 by Republicans, and one by Justin Amash of Michigan, who switched from Republican to Independent in July 2019.

Quinnipiac University and Emerson College: Mirror-image pollsters?

In three earlier posts—most recently here—I analyzed all polls conducted by Emerson College (“Emerson”) of 2020 presidential, senatorial and gubernatorial elections. I found that they had a clear bias towards the Republican candidate, on average, compared to all other polls of the same election.

As I continue to analyze polls of the presidential election between former Vice President Joseph R. Biden, Jr. and President Donald J. Trump, as well as this year’s 35 Senate and 11 gubernatorial elections, I have observed the opposite mathematical bias for Quinnipiac University (“Quinnipiac”) polls.

In fact, as we will see, the two polling organizations nearly mirror each other in their mathematical bias.

**********

Let us begin with the national race between Democrat Biden and Republican Trump. Using all 632 publicly-available polls released since January 1, 2019, I calculated a weighted-adjusted polling average (WAPA) of 8.3. That is, adjusting for time (with polls conducted after August 29 weighted and additional 1.5 times higher and polls conducted since September 29 3.0 times higher), pollster quality and partisan lean, Biden leads Trump by a little over eight points nationally. Emerson has an A- rating, and a historic lean of 0.3 points Democratic, while Quinnipiac has a B+ rating, and a historic lean of 0.2 points Democratic.

Table 1: State-level 2020 presidential election polling averages by Emerson College and Quinnipiac University compared to all other pollsters in the same state

StateQuinnipiac  Emerson 
 WAPAOtherDeltaWAPAOtherDelta
Arizona   4.62.72.0
California   29.328.60.8
Florida8.62.46.1   
Georgia5.70.25.5   
Iowa   -0.40.5-0.9
Kentucky-15.8-17.41.6   
Maine18.712.85.9   
Massachusetts   36.934.32.6
Michigan   11.07.43.7
Montana   -12.1-6.1-5.9
Nevada   -0.16.0-6.1
New Hampshire   7.88.1-0.3
North Carolina   0.92.1-1.2
Ohio1.40.41.0-0.30.5-0.8
Pennsylvania9.95.84.14.15.9-1.8
South Carolina-4.2-7.12.8   
Texas-0.9-1.60.7-1.1-1.60.4
Wisconsin   7.05.71.3
AVERAGE All States3.5  -0.5
AVERAGE Ohio, Pennsylvania, Texas1.9  -0.7

However, using the same calculation method—except for ignoring partisan lean—the 20 Emerson polls conducted by Emerson give Biden “only” a 3.7-point lead, which is 5.3 points more Republican than the average of the other 612 polls. At the same time, the 13 Quinnipiac polls give Biden a 9.9-point lead, which is 1.0 points more Democratic than the average of the other 619 polls. Put another way, Quinnipiac polls “see” a race that is 6.3 points more favorable to Biden than Emerson polls do, though both give Biden a solid lead.

While national polls are interesting—and plentiful—it is the Electoral College that determines who wins presidential elections. Table 1 compares state-level presidential polling averages by Emerson and Quinnipiac, in states where they have assessed the presidential election at least twice, to those calculated by all other pollsters in the state; positive values indicate a Democratic lead or pro-Democratic bias, and negative values indicate the opposite. Emerson has conducted one poll of the presidential election in Florida, Georgia, New Mexico, New Jersey and South Carolina; Quinnipiac has conducted one poll of the presidential election in Colorado, Connecticut, Iowa, Michigan, New Jersey, New York and Wisconsin.

In the 13 states where Emerson has assessed the 2020 presidential election, the average bias is only 0.5 points towards Trump, fully 4.8 points lower than its national bias toward Trump. By contrast, in the eight states where Quinnipiac has assessed this election—five of them in the south, the average bias is 3.5 points toward Biden, 2.5 points higher than its national bias toward Biden. The gap between the two polling organizations also narrows from 6.3 to 4.0 points at the state level.

Notably, while Quinnipiac has an average bias toward Biden in all eight states—ranging from more than five points in Georgia, Maine and Florida to around one point in Ohio and Texas—Emerson’s bias is evenly split across its 13 multiple-assessment states, ranging from nearly six points towards Trump in the western states of Montana and Nevada to between two and four points toward Biden in the disparate states of Arizona, Massachusetts (where Emerson College is located) and Michigan.

Curiously, Emerson and Quinnipiac have both assessed the 2020 presidential election in only three states—Ohio, Pennsylvania and Texas—multiple times since January 1, 2019. In these three states, the bias is relatively narrow: Trump +0.7 for Emerson and Biden +1.9 for Quinnipiac, for a gap of “only” 2.6 points.

**********

Turning to other statewide elections in 2020, Table 2 compares Senate election polling averages by Emerson and Quinnipiac, in states where they have conducted such polls at least twice since January 1, 2020, to those calculated by all other pollsters in the state. Emerson has conducted one Senate election poll in Georgia, assessing both seats on the ballot this year, as well as in Arizona, Iowa, Michigan, Minnesota, New Hampshire and New Jersey; Quinnipiac has conducted one Senate election poll in Iowa and Michigan. For the “jungle primary” in which Georgia Republican Senator Kelly Loeffler seeks reelection, I analyze the difference between the total percentage for all Democratic candidates and the total percentage for all Republican candidates.

Table 2: 2020 Senate and gubernatorial election polling averages by Emerson College and Quinnipiac University compared to all other pollsters in the same election

StateQuinnipiac  Emerson  
 WAPAOtherDeltaWAPAOtherDelta
Georgia–Loeffler-1.0-8.27.2   
Georgia–Perdue4.4-1.86.2   
Kentucky-9.5-9.3-0.2   
Maine9.14.74.5   
Montana   -8.4-0.8-7.6
North Carolina   2.65.1-2.4
South Carolina0.0-2.82.8   
Texas-7.0-6.5-0.5   
AVERAGE  3.3  -5.0

In a reverse of state-level presidential election polling, Quinnipiac has assessed six Senate elections multiple times—again focusing on the south, while Emerson has only assessed Senate elections in Montana and North Carolina more than once; no Senate election has been polled multiple times by both pollsters. Quinnipiac has an average 3.3-point bias toward Democratic Senate nominees, nearly identical to its 3.5-point state-level presidential election bias. By contrast, albeit only in two Republican-leaning states, Emerson has an average 5.0-point bias toward the Republican Senate nominees, nearly identical to their Trump bias nationally, and fully 4.5 points higher than their state-level bias toward Trump. Overall, and recognizing this is not an apples-to-apples comparison, Quinnipiac Senate election polls lean 8.3 points more Democratic than Emerson Senate election polls.

Quinnipiac has been especially Democratic-leaning in Georgia’s two Senate elections, with average pro-Democratic-nominee biases of 6.2 and 7.2 points, and in Maine, while they have shown minimal bias in Kentucky and Texas. As with the presidential election, meanwhile, Emerson has a whopping 7.6-point bias toward incumbent Montana Republican Senator Steve Daines in his race against Democratic Governor Steve Bullock; they are relatively closer to the mark in North Carolina, where Democrat Cal Cunningham has a small lead against incumbent Republican Thom Tillis.

Quinnipiac has conducted no gubernatorial election polls this year, while Emerson has conducted one poll in New Hampshire and multiple polls in Montana (6.5 points more Republican) and North Carolina (7.4 points more Republican), with a large average pro-Republican bias of 6.9 points!

**********

Table 3 lists the average partisan biases for Emerson and Quinnipiac for each set of races.

Table 3: Average partisan biases in Emerson College and Quinnipiac University polls across multiple election categories

StateQuinnipiacEmerson
Presidential–nationalDEM+1.0GOP+5.3
Presidential–stateDEM+3.5GOP+0.5
SenateDEM+3.3GOP+5.0
Governorn/aGOP+6.9
AVERAGEDEM+2.6GOP+4.4

While Quinnipiac has had only a relatively small pro-Biden bias in national polls, they have had an overall lean of 2.6 points Democratic across all elections they have assessed multiple times. By contrast, while Emerson has been very close to the all-polls average in their state-level presidential election polling, they have had an overall lean of 4.4 points Republican across all elections they have assessed multiple times. Overall, Quinnipiac has leaned fully 7.0 points more Democratic than Emerson has.

I will not attempt to “explain” these relative partisan biases, though they almost certainly result from some combination of how the demographic distribution of the likely 2020 electorate is modeled, the fact Quinnipiac shifted to “likely voters” models more recently than Emerson did, how hard they “push” initially undecided voters to choose one candidate, and the relative partisan leanings of demographic categories within their respective samples.

I will say, though, that the final Democratic-minus-Republican margin will almost certainly be very close to the midpoint of the two polling averages plus one point Democratic—at least in elections assessed at least once by Emerson and by Quinnipiac.

Until next time…please stay safe and healthy…and if you have not already voted, please do so as soon as you can!

Biden vs. Trump September 2020: A rising tide lifts more than enough boats

On November 3, 2020, the presidential election between incumbent Republican Donald J. Trump and his Democratic challenger, former Vice President Joseph R. Biden, Jr., will mark the end of a weeks-long electoral process. One month ago, I analyzed all publicly-available polls of the presidential election—nationally and by state, recognizing presidential elections are determined by the Electoral College—conducted since January 1, 2019.

Since then, Biden selected United States Senator Kamala Harris of California to be his vice-presidential running mate, as I anticipated, and both the Democrats and Republicans held mostly-virtual televised nomination conventions.

Did these events change the trajectory of this election?

The short answer is…no.

Table 1 lists the number of national polls assessing Biden vs. Trump conducted in 2019 and in each month of 2020; a handful of older polls were released since my last update. Sixty-five pollsters, with an average B- FiveThirtyEight pollster rating, have assessed the 2020 presidential election at least once since January 1, 2019; 45 of them (mean B-/B) have assessed the election more than once.

Table 1: Number of 2020 Monthly National Polls Assessing Biden vs. Trump

Month# Polls
All of 2019107
January 202020
February 202025
March 202035
April 202050
May 202048
June 202062
July 202051
August 202073[1]
TOTAL471

Fifteen pollsters (mean B-) account for 70% of these polls, as well as 68% of the 364 polls conducted in 2020:

  • YouGov (B-), 64 polls (49 in 2020)
  • Morning Consult (B/C), 48 polls (43 in 2020)
  • Ipsos (B-), 35 polls (28 in 2020)
  • HarrisX (C), 27 polls (18 in 2020)
  • Emerson College (B+), 19 polls (7 in 2020)
  • Fox News (A-), 18 polls (9 in 2020)
  • Change Research (C-), 16 polls (13 in 2020)
  • RMG Research (B/C), 15 polls (15 in 2020)
  • Data For Progress (B-), 14 polls (14 in 2020)
  • Optimus (B/C), 14 polls (13 in 2020)
  • IBD/TIPP (A/B), 14 polls (9 in 2020)
  • Rasmussen Reports/Pulse Opinion Research (C+), 13 polls (10 in 2020)
  • Quinnipiac University (B+), 11 polls (7 in 2020)
  • Zogby Interactive/JV Analytics (C+), 11 polls (6 in 2020)
  • CNN/SSRS (B/C), 10 polls (7 in 2020)

Figure 1 shows how Biden has fared monthly against Trump in 2020, using my weighted-adjusted polling averages (WAPA). I use pollster rating data to adjust for partisan lean (historic tendency to err more Democratic or Republican than other pollsters in analogous races) and quality. I weight more recent polls higher, using this ratio: number of days since January 1, 2019 divided by 673, the number of days between January 1, 2019 and November 3, 2020. I then average two versions of WAPA: one treating polls by the same pollster as statistically independent, and one treating all polls by the same pollster as a single, time-weighted value; differences between estimates are usually negligible.

Figure 1

Using all polls conducted since January 1, 2019, Biden leads Trump nationally by 7.5 percentage points (“points”). Biden’s margin rose from just over four percentage points in January and February, when he was fighting for the 2020 Democratic presidential nomination, to just under six points in the three months after Biden’s decisive win in the 2020 South Carolina Democratic presidential primary, to between eight and nine points since June 1, the day protesters were forcibly cleared from Lafayette Square so Trump could pose in front of St. John’s Episcopal Church holding a copy of the Bible.

**********

Again, though, presidential elections are fought across 50 states and the District of Columbia (“DC”), with the plurality winner in each state/DC winning every electoral vote (“EV”) from that state. Table 2 lists the number of polls within each state assessing Biden vs. Trump since January 1, 2019, plus that state’s 3W-RDM, my estimate of much more or less Democratic than the nation a state has voted over the last three presidential elections; eight states[2] and DC have not been polled.

Table 2: Number of state-level polls assessing Biden vs. Trump since January 1, 2019

State3W-RDMOverall2020
Michigan2.27561
Wisconsin0.77360
North Carolina-6.06956
Pennsylvania-0.46253
Florida-3.45345
Arizona-9.75244
Texas-15.34833
Georgia-9.63126
Ohio-5.82016
California23.22014
Iowa-4.72013
New Hampshire0.11510
Minnesota1.51413
Colorado2.21210
Virginia1.5128
Kentucky-28.7119
Maine5.9118
Montana-18.6109
South Carolina-15.7108
Missouri-15.997
Massachusetts22.187
Nevada2.084
New York21.677
Utah-33.176
Washington12.175
New Jersey12.066
Connecticut12.864
Alabama-28.455
Kansas-23.455
Mississippi-18.544
Oklahoma-38.144
Alaska-19.243
North Dakota-29.442
New Mexico6.533
Tennessee-25.833
Indiana-16.333
Maryland22.622
Delaware12.522
Arkansas-28.211
Hawaii34.311
Louisiana-22.211
West Virginia-35.511
TOTALD-6.1719582

Nineteen states have been polled at least 10 times since January 1, 2019, of which 14 have been polled at least 10 times in 2020. Four of the top five, along with suddenly-swing North Carolina, are the closest states won by Trump in 2016: Michigan, Wisconsin, Pennsylvania, Florida. Five other Republican-leaning states have been frequently polled: Arizona, Georgia, Texas, Ohio and Iowa, reflecting their status as ongoing or emerging battlegrounds. California, with 54 EV, rounds out the 11 states polled at least 20 times overall.

National averages still matter, though. Combined with 3W-RDM, they provide the “expected Democratic-minus-Republican margin” in each state in 2020, all else being equal. For example, North Carolina has recently been 6.0 points less Democratic than the nation as a whole. Adding that to Biden’s current national margin (-6.0 +7.5 = +1.5) suggests Biden is slightly favored to win North Carolina in 2020, based solely on its recent voting history. Indeed, Biden leads Trump by an adjusted mean of 1.9 points in 52 polls conducted in North Carolina. Table 3 lists every state’s expected value and WAPA.

Table 3: Expected and actual polling margins for Biden over Trump in each state in November 2020

State3W-RDMExpectedWAPAWAPA-Expected
DC82.089.5  
Hawaii34.341.829.1-12.7
Vermont27.735.2  
California23.232.728.4-2.3
Maryland22.630.124.6-5.5
Massachusetts22.129.634.34.7
New York21.629.125.9-3.2
Rhode Island18.025.5  
Illinois14.722.2  
Connecticut12.820.317.9-2.4
Delaware12.522.019.2-0.8
Washington12.119.624.65.1
New Jersey12.019.518.2-1.3
Oregon8.716.2  
New Mexico6.514.011.3-2.7
Maine5.913.410.4-3.0
Michigan2.29.77.0-2.7
Colorado2.29.712.12.5
Nevada2.09.53.8-5.7
Minnesota1.59.07.7-1.3
Virginia1.59.09.60.6
Wisconsin0.78.25.0-3.2
New Hampshire0.17.65.2-2.4
Pennsylvania-0.48.15.1-2.0
Florida-3.44.13.9-0.2
Iowa-4.72.8-1.6-4.4
Ohio-5.81.70.4-1.3
North Carolina-6.01.51.90.4
Georgia-9.6-2.1-0.71.3
Arizona-9.7-2.22.95.1
Texas-15.3-7.8-2.05.7
South Carolina-15.7-8.2-7.30.9
Missouri-15.9-8.4-6.32.2
Indiana-16.3-8.8-13.9-5.2
Mississippi-18.5-11.0-11.9-0.9
Montana-18.6-11.1-8.92.2
Alaska-19.2-11.7-4.37.3
Louisiana-22.2-14.7-11.23.5
Kansas-23.4-15.9-9.36.7
Nebraska-25.8-18.3  
South Dakota-25.8-18.3  
Tennessee-25.8-18.3-14.53.9
Arkansas-28.2-20.7-3.517.2
Alabama-28.4-20.9-18.22.7
Kentucky-28.7-21.2-17.04.2
North Dakota-29.4-21.9-20.41.5
Utah-33.1-25.6-13.112.5
Idaho-34.2-26.7  
West Virginia-35.5-28.0-34.3-6.3
Oklahoma-38.1-30.6-23.07.7
Wyoming-45.7-38.2  
AverageD-6.4Biden+1.5*Biden+2.1+0.7

* Only for the 42 states with both measures

The correlation between the expected margin and WAPA is +0.961, meaning polling matches expectations extremely well—as one increases or decreases, so does the other. Still, Biden is polling slightly ahead of those fundamentals, meaning state-level polling as a whole is even better for Biden than his excellent national polling; that said, the difference vanishes once you adjust for a state’s 2016 presidential election turnout.[3]

Biden is underperforming expectations in some states, most notably Hawaii—the birthplace of former President Barack Obama, artificially inflating Hawaii’s Democratic vote margin in 2008 and 2012. He is also underperforming in woefully-under-polled Nevada. Biden leads there by 3.8 points, nearly six points lower than the 9.5 points by which he “should” be leading. Biden is also underperforming expectations in very Democratic Maryland (-5.5) and Republican-leaning Iowa (-4.4). By the same token, Biden is overperforming in the traditionally Republican states of Arkansas, Utah, Oklahoma, Alaska, Texas and Kansas, as well as in reliably-Democratic Washington. There is a partisan split in Biden’s over-and under-performance: in states with 3W-RDM>-5.0, Biden is underperforming by 2.2 points, on average. In states with 3W-RDM≤5.0, Biden is overperforming by 3.4 points. Many grains of salt are in order here, though. In recent elections, “fundamentals” have missed the final margin by an absolute value average of 5.4 points.

Still, the close alignment between the two values allows us to combine them into a single estimate of Biden’s margin over Trump on November 3, 2020, assuming polls become more predictive as an election gets closer:

  1. Assign expected value and WAPA equal weight as of January 1, 2020.
  2. WAPA weight increases daily with proximity to November 3, 2020.

I also calculated how likely Biden is to win the EV from each state, assuming this likelihood is distributed normally:

  1. For expected margins, I use mean = -0.8 and standard error = 7.1[4]
  2. For WAPA, I use standard error = 3.0, roughly the margin of error in most quality polls; this is likely an over-estimate, as pooling reduces the standard error of the resulting polling average.
  3. Combined probability Biden wins a state’s EV calculated the same as for predicted final margin

While the means and standard errors I use are arguably arbitrary, albeit defensible, the final EV probabilities shown in Table 4 are in line with what other forecasters are saying.

Table 4: Estimated final state margins and probability of winning EV, Biden vs. Trump, November 2020

StateEVP(EV): ExpectedP(EV): WAPAP(EV): OverallPredicted Margin
DC3100.0% 100.0%89.5
Hawaii4100.0%100.0%100.0%31.1
Vermont3100.0% 100.0%35.2
California55100.0%100.0%100.0%28.7
Maryland10100.0%100.0%100.0%26.1
Massachusetts11100.0%100.0%100.0%33.8
New York29100.0%100.0%100.0%26.3
Rhode Island4100.0% 100.0%25.5
Illinois2099.9% 99.9%22.2
Connecticut799.7%100.0%99.9%19.1
Delaware399.7%100.0%100.0%19.6
Washington1299.6%100.0%99.9%22.1
New Jersey1499.6%100.0%99.9%18.8
Oregon798.5% 98.5%16.2
New Mexico596.8%100.0%99.3%11.9
Maine496.2%100.0%99.4%10.9
Michigan1689.6%99.0%98.0%7.3
Colorado989.5%100.0%98.8%11.9
Nevada688.9%89.5%89.4%4.4
Minnesota1087.6%99.5%98.3%7.9
Virginia1387.5%99.9%98.3%9.5
Wisconsin1085.2%95.2%94.2%5.3
New Hampshire483.1%96.0%94.3%5.5
Pennsylvania2081.3%95.5%94.0%5.3
Florida2968.0%90.5%88.1%4.0
Iowa661.0%29.6%34.4%-0.9
Ohio1854.9%55.7%55.6%0.6
North Carolina1554.0%73.6%71.6%1.9
Georgia1634.4%40.7%40.1%-0.8
Arizona1133.6%83.6%78.4%2.4
Texas3811.4%24.9%23.4%-2.7
South Carolina910.4%0.8%2.2%-7.4
Missouri109.7%1.8%2.7%-6.5
Indiana118.9%0.0%1.4%-13.1
Mississippi64.9%0.0%0.7%-11.8
Montana34.7%0.2%0.7%-9.1
Alaska34.0%7.4%6.9%-5.5
Louisiana81.5%0.0%0.2%-11.7
Kansas60.9%0.1%0.2%-12.6
Nebraska50.4% 0.4%-18.3
South Dakota30.4% 0.4%-18.3
Tennessee110.4%0.0%0.2%-16.4
Arkansas60.1%12.2%9.3%-12.1
Alabama90.1%0.0%0.0%-19.5
Kentucky80.1%0.0%0.1%-20.8
North Dakota30.1%0.0%0.0%-21.0
Utah60.0%0.0%0.0%-15.1
Idaho40.0% 0.0%-26.7
West Virginia50.0%0.0%0.0%-31.2
Oklahoma70.0%0.0%0.0%-26.8
Wyoming30.0% 0.0%-38.2
  • He is at least an 88.1% favorite in enough states—and by margins of at least four points—to earn him 308 EV, or 307 depending on what happens in Maine, which, along with Nebraska, allocates two EV to the statewide winner and one each to the winner of its Congressional districts. Moreover, Biden could lose Florida (+4.0, 88.1%), Nevada (+4.4, 89.4%) and one EV in Maine and still win 272 EV, two more than he needs.
  • He is a 70-75% favorite to win in Arizona (+2.4) and North Carolina (+1.9), for an additional 26 EV, increasing Biden’s total to 333/334 EV.
  • The 34 combined EV of Ohio (+0.6) and Georgia (-0.8) are essentially toss-ups, meaning Biden has a roughly 73% chance to win at least one of them, putting him somewhere between 349 and 352 EV, with a maximum of 368 EV (or 369 with one EV in Nebraska).

Three months before Election Day 2020, and with every caveat about voting during a pandemic, Joe Biden is the strong favorite to be elected the 46th president of the United States.

Plus, it might take only a sharp break by undecided voters and a modest polling error for Biden to win the 44 combined EV of Iowa (-0.9) and the ultimate prize—Texas (-2.7). Thus, while something in the low-to-mid 300’s currently appears the most likely EV total for Biden, 413 EV cannot be discounted.

Using the simplistic—perhaps even simple-minded—method of multiplying Biden’s probability of winning each state by its EV and summing yields a “projected” EV total of 349.2, essentially adding Michigan, Pennsylvania, Wisconsin, Florida, Arizona, North Carolina, and one of Ohio/Georgia to the states 2016 Democratic presidential nominee Hillary Clinton won.

Biden’s lead looks even more robust when you make either of two historically-valid assumptions:

Polls systematically overestimate Biden’s margins by 3.0 points.

In this scenario, Biden’s projected EV drops to 301.2, still 31.2 more than required. He would be favored at least 80% to win in enough states to win 239 EV, though he would also be favored by at least 74.6% in three states totaling 34 EV, getting him to 273 EV. Thus, even if Biden “only” wins the national popular vote by 4.2, he would likely still prevail, though the decisive states—some combination of New Hampshire, Pennsylvania and Wisconsin—would be decided by narrow margins, with all votes possibly not counted for a week or more.

Polls systematically underestimate Biden’s margins by 3.0 points.

In this scenario, Biden’s projected EV are a landslide-level 389.4, 119.4 more than necessary. He would be favored at least 80% to win enough states to earn 352 EV, while being a 74.9% favorite in Georgia, for a total of 368 EV. He would also be a 69.1% favorite in Iowa, with Texas now a toss-up at 58.1%–and a projected Biden victory of 0.3 points! Based on only one poll, Biden would have a 33.2% chance of winning Arkansas’ 6 EV, plus a 14.3% chance of winning Missouri’s 10 EV and a 11.5% chance of winning Alaska’s 3 EV. The last presidential candidate to come close to 433 EV was Republican George H. W. Bush, who won 426 EV in 1988.

**********

To maximize the number of polls available for analysis, I use all polls going back to January 1, 2019; I also use a straightforward time-weighting method: increasing the weight of a poll by 1/673 = 0.0015 every day since then.

To test the validity of this weighting system, I re-estimated every WAPA and probability using the following time-weighting system, based on the key dates of February 29 and June 1 described above:

  1. Only use polls whose field date midpoint is January 1, 2020 or later (that is, date weight for 2019 polls = 0)
  2. Weight polls released in January and February 2020 as before
  3. Weight polls released in March, April and May 2020 twice as much as before
  4. Weight polls released in June, July and August 2020 three times as much as before
  5. Weight polls in September and October 2020 four times as much as before

This system gives vastly more weight to the most recent polls and correspondingly much lower weight to earlier polls.

As one would expect from Figure 1, Biden’s national lead jumps to 8.0 points using this time-weighting method, though the projected EV total barely increases to 351.4, with only minor changes in the probability Biden wins any given state: Nevada (92.3% and Florida (90.7%) rise slightly, while Ohio (52.2%) and Georgia (40.1%) decline slightly.

However, state-level changes in WAPA are very telling, as Table 5 reveals:

Table 5: 2020 Polling Margins, Biden vs Trump, Using Two Time-Weighting Methods

StateWAPA Original Time-WeightWAPA Recent Time-WeightDelta
Hawaii29.1029.100.00
California28.3729.911.55
Maryland24.6024.49-0.11
Massachusetts34.3034.300.00
New York25.8927.031.15
Connecticut17.9118.620.71
Delaware19.1720.040.87
Washington24.6526.131.49
New Jersey18.2018.220.02
New Mexico11.3011.770.47
Maine10.4010.710.31
Michigan7.027.570.55
Colorado12.1513.171.02
Nevada3.764.340.58
Minnesota7.738.340.60
Virginia9.5610.921.36
Wisconsin5.005.930.94
New Hampshire5.246.100.86
Pennsylvania5.075.200.13
Florida3.934.530.60
Iowa-1.61-1.080.53
Ohio0.430.11-0.32
North Carolina1.892.150.26
Georgia-0.70-0.87-0.17
Arizona2.933.010.08
Texas-2.03-1.920.10
South Carolina-7.25-6.620.63
Missouri-6.28-5.710.57
Indiana-13.92-14.48-0.56
Mississippi-11.92-11.500.41
Montana-8.86-8.360.51
Alaska-4.33-4.010.32
Louisiana-11.19-11.190.00
Kansas-9.27-9.030.24
Tennessee-14.47-14.280.19
Arkansas-3.50-3.500.00
Alabama-18.19-17.950.24
Kentucky-17.05-17.82-0.77
North Dakota-20.42-19.131.28
Utah-13.11-13.88-0.77
West Virginia-34.30-34.300.00
Oklahoma-22.96-22.030.93
AverageBiden+2.08Biden+2.480.40

Extending WAPA to two decimal places, Biden’s national lead increases by 0.46 points, from 7.50 to 7.97. However, rather than Biden increasing his lead by four or five points in some states, say, while decreasing his lead by three or four points in other states, only five states saw a decline in Biden’s average polling margin—Maryland, Georgia, Ohio, Kentucky and Utah; no decline was less than -0.77. And in only six states—Colorado, New York, North Dakota, Virginia, Washington and California—did Biden’s average polling margin increase by more than 1.00 points, with a maximum of 1.55. The standard deviation of the average change in Biden’s margin is only 0.55, showing how tightly bunched around the mean of 0.40 points these shifts are.

In other words, when switching to a time-weighting method which gives vastly more weight to polls released over the preceding three months while eliminating 2019 polls entirely, Biden saw his lead either not change or increase by up to 0.94 in 31 of 42 states. This remarkably consistent change should alleviate fears that Biden will win the popular vote by four or five points, yet still lose the Electoral College because he won even more votes than Clinton in safe states like California and New York while narrowing the 2016 margins in states like Georgia, Ohio and Texas without actually winning any of their EV. Instead, as Biden’s national margin increases, his lead in nearly every state—including nearly every swing state—increases correspondingly. Put differently, the same groups of voters fueling increases in Biden’s vote total in one set of states are also fuel increases in states like Michigan, Wisconsin, Pennsylvania and Florida—and perhaps Arizona and North Carolina as well.

Figure 2 makes this same point in a different way. It compares current Biden WAPA to simply increasing every 2016 Democratic margin increased by 5.4 points, the difference between Biden’s current national polling lead and the 2.1 points by which Clinton won in 2016.

Figure 2

Figure 2 perfectly illustrates the adage “A rising tide lifts all boats,” or nearly all, anyway. Biden’s current state-level polling averages—as I calculate them—are astonishingly close to how you would expect him to fare in each state given a 5.4-point increase in the national Democratic margin.

**********

None of this is to say Biden is guaranteed to be elected president of the United States on in two months. There are worrisome signs this year’s elections will not be conducted as efficiently and transparently as they could be. Delays in mail delivery—allegedly orchestrated by a newly-confirmed Postmaster General—could leave millions of votes uncounted because they did not arrive by November 3. Moreover, while Biden’s national polling lead has consistently ranged between four and 10 points over the last 20 months, a late-recovering economy or last-minute “October surprise” could erase this lead.

All that being said, however, unlike Clinton in 2016, Biden has a sufficiently-wide range of paths to 270 EV that I estimate he is at least an 80% favorite to be elected president of the United States on November 3, 2020—or whenever ballots are ultimately counted.

Until next time…please stay safe and healthy…


[1] Includes one Redfield & Wilton Strategies poll conducted August 31 to September 1

[2] DC, Vermont, Rhode Island, Illinois, Oregon, Nebraska, South Dakota, Idaho, Wyoming

[3] That said, this does not account for mid-to-large states like Oregon and Illinois where Biden is expected to win by double-digit margins.

[4] The former value is the mean arithmetic difference between “expected” and actual D-R margins across 153 state-level contests in 2008, 2012 and 2016, while the latter value is the standard deviation of these values. I recognize this is not a standard error. However, using the value 13.6—the range of values covering 95% of all values divided by 1.96, the final EV projection changes by only 1.0.

An update on Emerson College polling

In two essays I published in November 2019 (here and here), I addressed differences between polling results obtained by highly-respected Emerson College and those from other pollsters in the same race. Emerson College has an A- rating from FiveThirtyEight.com, and their analysts calculate that, on average, Emerson College polls are about 0.3 points more Democratic than other pollsters in the same elections.

Emerson College recently released new national and state-level polls, renewing interest in their work. I therefore decided to update my assessment by comparing Emerson College polls to those released by other pollsters in the same races, using all publicly-available polls released since January 1, 2019.

I Voted sticker

I begin with national polls of the 2020 presidential election between Democrat Joseph R. Biden, Jr., the former Vice President, and Republican incumbent Donald J. Trump. The 18 polls Emerson College has released of this race, weighted by time, average to a Biden lead of 3.6 percentage points (“points”) over Trump; this increases to 3.9 if I do not adjust for Emerson’s calculated Democratic lean. By comparison, the other 370 polls of this race—using my weighted-adjusted polling average (WAPA)—give Biden a 7.4-point lead, with essentially no partisan lean. That is, relative to other polls assessing Biden vs. Trump nationally, Emerson College polls are 3.5-3.8 points more Republican.

Table 1 repeats this comparison for every state in which Emerson College has polled the presidential election at least twice since January 1, 2019; the month of the most recent Emerson College poll is also listed. States are sorted by number of Emerson College polls and poll recency.

Table 1: Comparing state-level WAPA, Biden vs. Trump, Emerson College to all other pollsters

State # Emerson Polls Emerson College All Other Pollsters Emerson minus Other
Iowa 4

(12/2019)

-1.8 -2.2 D+0.4
Texas 3

(5/2020)

-1.4 -2.1 D+0.7
New Hampshire 3

(11/2019)

6.7 4.7 D+2.0
Massachusetts 2

(5/2020)

34.4 30.6 D+3.8
Ohio 2

(5/2020)

-0.4 1.0 D-1.4
California 2

(5/2020)

29.3 27.6 D+1.9
Michigan 2

(11/2019)

11.0 7.0 D+4.0
Nevada 2

(11/2019)

-0.4 +4.6 D-4.2
TOTAL/AVERAGE 20 9.7 8.9 D+0.8
Weighted by # polls 20 7.8 7.0 D+0.8

These eight states tell a very different story. On average, these polls show an average 0.8 points more support for Biden than all other polls in these states, irrespective of the number of polls Emerson College has conducted in that state. One major caveat is that Emerson College has yet to conduct a presidential horse race poll in calendar year 2020 in four of these states. If anything, though, state-level Emerson College polls of Biden vs. Trump have become more Democratic leaning: the four states with no calendar year 2020 polls show a Democratic lean of 0.45, broadly similar to the FiveThirtyEight estimate of 0.3, compared to 1.25 for the four states with a May 2020 Emerson College poll. Remove Massachusetts (home to Emerson College), however, and the difference vanishes.

I will post my assessment of polling in 2020 election for United States Senate (“Senate”) and governor around Labor Day; this assessment will only use polls released since January 1, 2020. At this point, though, Emerson College has released only two polls, in total, of any of this year’s 35 Senate and 11 gubernatorial elections. Both were conducted in Montana with 531 likely voters between July 31 and August 2, 2020. In the Senate race, Democratic Governor Steve Bullock trails incumbent Republican Steve Daines 44-50. The other six polls of this election give Bullock a narrow 0.9-point lead, for a pro-Republican lean of 6.9 points. Similarly, in the open governor’s race, Democratic Lieutenant Governor Mike Cooney trails Republican United States House of Representatives member Greg Gianforte 41 to 50; three other polls have Cooney down by 5.2 points, for a pro-Republican lean of 3.8 points. With all necessary caveats for small numbers, the average pro-Republican bias in these two polls is 5.4.

Curiously, while Emerson College presidential election polls force respondents to choose only between Biden and Trump—meaning their vote shares sum to 100%–in the two recent Montana polls, an average 7.5% of respondents chose an option besides the named Democrat and named Republican.

Still, this does not explain why, at the national level, Emerson College 2020 presidential election polls lean 3.5-3.8 points more Republican, while the sparse state-level polling leans 0.8 points more Democratic. Splitting the difference implies an Emerson College pro-Republican lean of about 1.4 points, but I would like to see far more state-level polling to have any confidence in that value.

Until next time…please stay safe and healthy…

Emerson College polls: Post, and ye shall receive

Five days ago, I wrote about the sharp Republican shift since September 1, 2019 in polls conducted by Emerson College of hypothetical 2020 presidential election matchups between President Donald J. Trump and possible Democratic nominees. Emerson College is a high-quality pollster, rated A- and showing no partisan skew in recent elections according to FiveThirtyEight.com’s pollster ratings. What distinguishes Emerson College from other similarly high-quality pollsters assessing these matchups, though, is that the results they release have 0% choosing undecided or an unnamed candidate; this is likely, at least in part, because approximately 70% of Emerson College polls are conducted using automated voice response (aka “robopolling”).

[Eds. note: See August 2020 update]

I Voted sticker

When I first wrote in June 2019 about Emerson College polls of these hypothetical matchups, it was to compare their results with those of HarrisX, a C+ pollster with a strong Republican bias of 1.3 percentage points (“points”). I calculated then that “[HarrisX”] polls have an average of 28.0% undecided between the named Democrat and Trump (or would choose a third-party candidate); I estimate these voters would break roughly 7-4 in favor of the Democratic nominee.” That rises to 2-1 if polls assessing former Vice President Joe Biden or United States Senator (“Senator”) from Vermont Bernie Sanders are excluded.

However, as though in response to my November 25, 2019 post, Emerson College just released results of hypothetical 2020 presidential election matchups in New Hampshire. They assessed five possible Democratic nominees: Biden, Massachusetts Senator Elizabeth Warren, Sanders, South Bend, IN Mayor Pete Buttigieg, and entrepreneur Andrew Yang; they surveyed 637 registered voters between November 22 and November 26,  2019.

Actually, the FiveThirtyEight.com poll tracker lists two sets of results: one with 0% “undecided/other” and one with between nine and 12% “undecided/other” (mean=10.8%); note that the former results are based upon 547 voters, 90 fewer than the latter results. We can thus directly assess how those roughly 1 in 9 voters would vote if “forced” to choose; results may be found in Table 1.

Table 1: Comparing Emerson College Polls of Hypothetical Match-ups Between President Trump and Specific 2020 Democratic Presidential Nominees in New Hampshire

Democrat With Undec/Other Without Undec/Other %Undec/Other %Trump by Undec/Other
Biden D+4 D+5 12% 45.8%
Warren D+4 D-2 10% 80.0%
Sanders D+7 D+3 9% 72.2%
Buttigieg D+8 D+6 11% 59.1%
Yang D+5 D-2 12% 79.2%
AVERAGE D+5.6 D+2.0 10.8% 66.7%

Poll results with a non-zero percentage choosing “undecided/other” show the hypothetical 2020 Democratic presidential nominee beating Trump in New Hampshire by an average of 5.6 points, slightly higher than the 4.6 points I calculate using all polls released since January 1, 2019. Moreover, all five Democrats lead Trump by between four and eight points. Once “undecided/other” voters are forced to choose, however, that average drops sharply to 2.0 points; Warren and Yang now trail Trump. That is because, on average, “undecided/other” voters choose Trump 2-1; this masks a curious divide, though: when offered Biden or Buttigieg, the split is nearly even, but when offered Warren, Sanders or Yang, the split is closer to 3-1.

While not strictly an apples-to-apples comparison, “undecided/other” voters breaking 2-1 for Trump is an exact reversal of my June analysis, when analogous voters broke nearly 2-1 against Trump. This reversal feels…implausible…if only because for years, the general rule of thumb was that these voters would ultimately break toward the non-incumbent, perhaps by as much as 2-1, as the incumbent was such a well-known quantity that opinions were basically fixed: a voter already knew if they approved of her/him or not. Still, that “rule” no longer seems to apply, so the best thing to do is to continue to aggregate all polls as best we can.

At the same time, the Emerson College results, when applied to other recent polls of hypothetical 2020 presidential election match-ups, are…odd. For example, consider this comparison of national SurveyUSA (A rating, D+0.1; 3,850 registered voters [RV]. November 20-21, 2019) and Emerson College polls (1,092 RV, November 17-20, 2019):

Table 2: Comparing National November 2019 SurveyUSA and Emerson College Polls of Hypothetical Match-ups Between President Trump and Specific 2020 Democratic Presidential Nominee

Democrat SurveyUSA Emerson College SurveyUSA

%Undec/Other

%Trump by Undec/Other
 

Biden

52%

D+13

49.5%

D-1

9% 155.6%
 

Warren

49%

D+7

50%

D+0

9% 88.9%
 

Sanders

52%

D+12

50.5%

D+1

8% 137.5%
 

Buttigieg

48%

D+7

48%

D-4

11% 100.0%
 

AVERAGE

50.2%

D+9.8

49.8%

D-0.5

9.2% 120.5%

In the four SurveyUSA polls, the hypothetical 2020 Democratic presidential nominee tops 50% on average, with a nearly-10-point lead over Trump and only about 1 in 11 choosing “undecided/other.” By contrast, the four Emerson College polls—with every respondent forced to choose either Trump or his Democratic opponent—show a dead-even race, with Trump edging the Democrat by 0.5 points, on average. But for these two sets of polls to exist simultaneously, not only would every single (or, in the case of Warren, 89%) “undecided/other” voter have to break for Trump, but a small proportion of Democratic-leaning voters would also have to switch to Trump as well.

Which, again, seems…implausible.

As always, caveat emptor.

Until next time…

About those recent Emerson College polls…

I first wrote about Emerson College polls here, using the fact their polls of hypothetical 2020 matchups between a Democrat and President Donald J. Trump force respondents to choose a candidate (i.e., have 0% “other/undecided”) to assess Harris X polls, which often have very high proportions “other/undecided.” At the time, I concluded “other/undecided” Harris X polls respondents likely would vote for the 2020 Democrat presidential nominee roughly 2-1.

[Eds. note: I have since updated this post here and here]

When I wrote that post in June, margins reported by Emerson College polls were broadly in line with those reported by other, non-Harris-X polls of these hypothetical matchups. Since the end of August, however, they have taken a sharply Republican turn compared both to previous Emerson College polls and to all other polls, based upon analyses of my WAPA (weighted-adjusted polling average).

Table 1 clearly demonstrates this pro-Republican shift.

Table 1: Polling margin for specified 2020 Democratic presidential nominees over President Trump, Emerson College vs. All Other Polls and January-August 2019 vs. September-November 2019

Biden–National
Emerson All Other
Jan-Aug 7.24 8.35
Sep-Nov 0.29 8.03 Difference
-6.95 -0.32 -6.63
Warren–National
Emerson All Other
Jan-Aug 1.52 1.82
Sep-Nov 1.29 4.66 Difference
-0.23 2.84 -3.07
Sanders–National
Emerson All Other
Jan-Aug 4.28 4.99
Sep-Nov 0.12 5.10 Difference
-4.16 0.11 -4.27

While the post-August-2019 Republican lean of Emerson College polls for the three current polling leaders for the 2020 Democratic presidential nomination—former Vice President Joe Biden, Massachusetts United States Senator (“Senator”) Elizabeth Warren and Vermont Senator Bernie Sanders—is apparent, it is most notable for Biden. Thus, while in all other polls, Biden’s WAPA declined modestly from 8.35 to 8.03, in the Emerson College polls it dropped from 7.24 to essentially even! Put another way, there was a 6.63 shift Republican in Emerson College polls for Biden starting around September 1, 2019. The shift was similar for Sanders—4.16 more Republican in Emerson College polls and no meaningful shift in other polls. Warren, meanwhile, had a sharp increase in her WAPA—from 1.82 to 4.66—in non-Emerson-College polls after September 1, 2019, while dropping slightly in the Emerson College polls. Overall, across all three candidates, there was a 4.66 pro-Republican shift in Emerson College polls after September 1, 2019.[1]

I Voted sticker

**********

Before addressing what could have caused this shift, let us address what did NOT cause this shift. On November 5, 2019, FiveThirtyEight.com updated its pollster ratings, which I use in calculating WAPA. Emerson College was upgraded from B+ to A-, making it one of the highest-quality pollsters regularly assessing the 2020 presidential election. Moreover, Emerson College’s “mean-reverted bias”—how much more or less Democratic its average polling margins are compared to other pollsters in the same (already-concluded) election—barely changed, shifting from D+0.1 to R+0.0 (which I code as D-0.025). Finally, mean sample size of Emerson College polls—still of registered voters—dropped only slightly after September 1, from 1,120 to 1,043.

But what about other pollsters? If anything, the mix of pollsters assessing the 2020 presidential election improved after September 1, 2019, from B-/B to B, albeit with a slightly more Democratic skew (D+0.2 to D+0.3).

Moreover, here are the averages of unskewed pollster-average margins for Biden, Warren and Sanders across all pollsters with at least a B+ rating,[2] before and after September 1, 2019:

  • Biden: 9.3 percentage points (“points”) to 10.8 points
  • Warren: 2.4 points to 6.6 points
  • Sanders: 5.8 points to 7.5 points

Clearly, both the more recent Emerson College polls and the lower-rated pollsters are finding much closer races between each of these three candidates and Trump than are the higher-rated pollsters. And while I removed the skew from these margins, the higher-rated pollsters have a mean skew of D+0.5—so take these averages with a modicum of salt.

Nonetheless, polls conducted by pollsters similar in quality to Emerson College—albeit with some small percentage of “undecided/other” voters—show the three leading Democratic candidates increasing their hypothesized margins against Trump by 1.5 to 4.2 points after September 1, 2019, while the Emerson College polls show declines in support from 0.2 to 7.0 points.

So what gives?

There are two broad categories of possible, non-mutually-exclusive explanations.

  1. Starting September 1, Emerson College pollsters adjusted how they weigh their samples by various demographic factors, such that their polls skew sharply more Republican relative to other high-quality pollsters.
  2. Emerson College pollsters, perhaps because they force respondents to make a choice, are capturing a genuine pro-Republican shift in the electorate other high-quality pollsters are missing.

Taking each possible explanation in turn…

Emerson’s sampling methodology did not change after September 1, remaining a combination of Interactive Voice Response (i.e., “robocalls”) of landlines—but not cellphones—and an online voter panel. However, their sample weighting, based upon 2016 turnout, did change subtly—from age, region, income, and education in their August 2019 national polls to age, [interview] mode, party registration, ethnicity and region in their November 2019 national polls.

Given how strongly education is now associated with partisanship, especially among white voters, that could account for at least some of the difference. I am perplexed, however, how interviewing mode is associated with turnout—other than landline-users tending to lean Republican (older, whiter, less urban). For that matter, I am not clear why a poll of registered voters would adjust for turnout at all. Adjust for the relative proportions of these groups in the universe of registered voters, sure—but adjustment for participation rates (i.e., turnout) of various groups in 2016 seems more appropriate for a likely voters model. Still, without seeing the raw data, I will not even speculate how these changes in sample weighting would affect publicly-released polling margins.

There is, meanwhile, an argument to be made that the universe of “decided” voters has drifted Republican in recent months, especially with the announcement by Speaker of the United States House of Representatives (“House”) Nancy Pelosi on September 24, 2019 that the House would formally begin an impeachment inquiry into Trump. While polling at first showed approval for this action higher than disapproval, the difference is closer to even now as House Republicans have rallied behind a president of their party.

Also, while Democrats fared well in recent governor’s races in Kentucky, Louisiana and Mississippi, the final actual margins were an average 2.8 points more Republicans than the final WAPA, perhaps reflecting a substantial Republican bias among “other/undecided” voters—essentially what the Emerson College polls might show. Moreover, the A+-rated Siena College/New York Times Upshot (R+0.3), recently released a set of state-level polls that also show leading Democrats faring less well against Trump than the consensus of other pollsters in those states.

While some combination of these two possible explanations—a change in sample weighting and an actual pro-Republican shift in the electorate—probably accounts for the clear shift towards Trump in recent Emerson College polls, it is entirely possible their last three national polls are showing these shifts purely by chance; even the best sampling strategy will be well wide of the mark at times (this is, in fact, the basic logic behind polling aggregation). However, three such wide sampling “misses” in the same direction and of the same size are extremely unlikely, though far from impossible.

In the end, the best those of us who track election polling can do is throw every publicly-available poll into the analysis, weighting, averaging and adjusting as best we can—all the while remembering that even the best estimates are just that, estimates.

Until next time…

[1] Values were similar for South Bend, Indiana Mayor Pete Buttigieg and California Senator Kamala Harris, but with each having only one post-August-2019 Emerson College poll assessing support versus President Trump, I excluded their polls from the analysis.

[2] IBD/TIPP, CNN/SSRS, Fox News, NBC News/Wall Street Journal, ABC News/Washington Post, SurveyUSA, Quinnipiac University. There is no post-August-2019 NBCNWSJ polling for Sanders.