A note to readers: I have temporarily stopped writing “dispatches” about how my wife Nell, our two daughters and I cope with social distancing and the closure of Massachusetts schools through the end of the 2019-20 school year because they started to feel repetitive. When and if that changes, I will resume dispatching.
As I write this, it is exactly six months until the 2020 United States (U.S.) presidential election, which will conclude on November 3, 2020. On April 8, 2020, U.S. Senator from Vermont Bernie Sanders announced he was suspending his campaign for the 2020 Democratic presidential nomination, making former Vice President Joseph R. Biden, Jr. the presumptive nominee against incumbent Republican president Donald J. Trump.
Using all publicly-available polls of the presidential election—both nationally and at the state level, recognizing presidential elections are determined by the Electoral College—conducted since January 1, 2019, I have been tracking the relative performance of contenders for the 2020 Democratic nomination against Trump. When given the choice, I used polls of likely voters over those of registered voters, and the latter over polls of adults only; I also used polls including such possible third-party candidates as former Starbucks CEO Howard Schultz and U.S. House of Representatives Member Justin Amash of Michigan. Table 1 lists the number of national polls conducted each month for both candidates based upon the midpoint of the poll’s field dates; some polls were actually conducted in two months.
Table 1: Number of National Polls Assessing Hypothetical 2020 Match-ups Between Biden/Sanders and Trump by Month
Just seven of 41 total pollsters (average grade: B-/B) account for 54% of Biden versus Trump polls; the values are similar for Sanders:
- IBD/TIPP (A/B), 10 polls
- Fox News (A-), 13 polls
- Harris X (C+), 13 polls
- Emerson College (B+), 18 polls
- Ipsos (B-), 18 polls
- Morning Consult (B/C), 22 polls
- YouGov (B-), 36 polls
Figure 1, meanwhile, shows how Biden and Sanders fared monthly against the president, using my weighted-adjusted polling averages, or WAPA. Basically, I use data published by FiveThirtyEight.com to adjust each poll for partisan lean (tendency of a pollster to err more Democratic or Republican than other pollsters in analogous races) and overall quality (using the letter grade assigned by FiveThirtyEight.com). I also weight more recent polls—again using field midpoint—higher, using the ratio of the number of days since January 1, 2019 and the total number of days between January 1, 2019 and November 3, 2020. Finally, I average two different versions of WAPA: one treating polls by the same pollsters as statistically independent values, and one which treats all polls by the same pollster as a single value; differences between estimates are generally negligible.
Figure 1: Monthly weighted-adjusted average margins for Biden and Sanders versus Trump since January 2019
Only one national poll assessing hypothetical matchups between Biden or Sanders and Trump was conducted in January 2019, so I combined them with the four and three, respectively, from February 2019 to generate Figure 1. Biden and Sanders have consistently led Trump in head-to-head matchups, never dropping below Sanders’ 2.0 percentage point (“points”) lead in December 2019. Through September 2019, Biden’s margin was typically three-to-four points higher, though Sanders still led Trump by 4.3 points on average, versus 7.8 points for Biden. From October 2019 through February 2020, though, the two men fared equally well versus Trump, with Biden ahead an average 5.4 points and Sanders ahead 4.9 points. Once Biden’s nomination began to become clear in March 2020, however, Biden again began to fare better versus Trump than Sanders, averaging a 5.7-point-lead to Sanders’ 3.4-point lead. Overall, Biden has a 6.1-point lead over Trump, not meaningfully different than his lead over the last two months; Sanders exited the race with an overall national lead of 4.3 points versus Trump, though that lead had begun to drop slightly over the last two months.
Again, however, presidential elections are actually fought across all 50 states and the District of Columbia (“DC”), with the plurality winner in each state/DC winning every electoral vote (“EV”) from that state.
To that end, Table 2 lists the number of polls conducted within each state since January 1, 2019 of hypothetical matchups between Biden/Sanders and Trump, plus that state’s 3W-RDM, an estimate of much more or less Democratic than the nation a state tends to vote; 11 states and DC have not yet been polled.
Table 2: Number of state-level polls assessing hypothetical 2020 matchups between Biden/Sanders and Trump since January 1, 2019
It is not surprising that eight of the 14 most-polled states thus far are “swing” states, those with 3W-RDM between -5.0 and +5.0, including the four closest states won by Trump in 2016: Florida (19 Biden, 11 Sanders), Pennsylvania (23, 17), Wisconsin (30,26) and Michigan (33,23). In fact, the Pearson correlation between the absolute value of a state’s 3W-RDM and the number of times it has been polled for the 2020 presidential election is -0.47 for Biden and -0.48 for Sanders, meaning the closer a state is to the national average (i.e., a pure toss-up in a dead-even national race), the more often it has been polled. Also highly-polled are large states like California and Texas, red-drifting states like Ohio and Iowa, and emerging Democratic opportunities like Arizona, Georgia and North Carolina.
While U.S. presidential elections are decided on a state-by-state basis, though, national averages are still important. Combined with 3W-RDM, they provide the “expected Democratic-minus-Republican margin” in each state in 2020, all else being equal. Comparing polling averages to this expected value tells us where Biden may currently be under- or over-performing, or which states have drifted Democratic or Republican since 2016.
For example, Biden leads Trump overall by 6.1 points. North Carolina has recently been about 6.0 points less Democratic than the nation as a whole. Adding those two values together (6.1 – 6.0 = +0.1) yields an expected photo-finish in North Carolina in 2020. However, Biden leads Trump by a mean 2.2 points in 23 polls thus far in North Carolina, meaning Biden is “outperforming” expectations there by about 2.1 points.
This could mean any or all of three things:
- WAPA is the more accurate reflection of the November election and either
- North Carolina has drifted about two points toward the Democrats since 2016, or
- The true “expected value” is somewhere between Trump winning by 5.3 points and Biden winning by 5.5 points, based upon an average 3W-RDM error margin of 5.4 points in recent elections.
- The “expected” value is the more accurate reflection, and Republican-leaning voters will drift back toward Trump over the next six months, making North Carolina nail-bitingly close on election day.
Table 3 lists every state’s expected value and WAPA; for ease of presentation, I include Biden-Trump values only.
Table 3: Expected and actual polling margins for Biden over Trump in each state in November 2020
* Only for the 39 states with both measures
The correlation between the expected margin and WAPA is a very-reassuring +0.96, meaning the polling is broadly in line with the underlying “fundamentals” of the election. Still, Biden is polling ahead of those fundamentals by an average of about one percentage point, meaning the state-level polling as a whole is even better for Biden than his already-solid national polling.
Nonetheless, there are clearly states where Biden is underperforming expectations, including the vital and heavily-polled state of Wisconsin. While Biden leads there by about 1.7 points overall, he “should” be ahead there by about 6.8 points. Moreover, he is trailing by about 3.5 points in nearby Iowa, even though Biden “should” be ahead by about 1.4 points. And while Biden leads Trump by about 3.5 points in Nevada, that is 4.6 points below what the fundamentals suggest.
The story is similar, but more narrowly so, in the key states of Michigan, Pennsylvania, New Hampshire and Florida: Biden leads Trump in these states by an average of 4.1 points, though he “should” lead by an average of 5.8 points, a mean “underperformance” of 1.7 points.
Moreover, there appears to be something of a partisan split in Biden’s over-and under-performance: in the 10 states with both measures and 3W-RDM≥5.0, Biden is underperforming by 0.3 points, on average, though once you remove the single poll of Massachusetts, that jumps to -1.6 points. At the same time, in the analogous 20 Republican states with 3W-RDM≤5.0, Biden is overperforming by 3.2 points, though that drops to 2.6 with the massive outlier of Utah removed.
Let me again stress, however, that there is a lot of “wobble” in the “expected margins,” as well as in the polling averages—especially given that most states have seen very little recent polling. All of this “over- and underperforming” may simply be statistical noise, as we try to read too much into highly stochastic data.
Still, the two values are sufficiently closely aligned to combine them into a single, six-months-out estimate of Biden’s margin over Trump on November 3, 2020, based upon the assumption polls become more predictive as an election gets closer:
- Arbitrarily assign expected value and WAPA equal weight as of January 1, 2020.
- If the most recent poll in a state was conducted more than 100 days prior to January 1, 2020, WAPA is weighted just 10%. This only applies to Massachusetts, Alaska and Kentucky, with Minnesota the only other state whose most recent poll was conducted in 2019.
- WAPA weight increases, by day, with proximity to November 3, 2020.
At the same time, I introduced a probabilistic element into these estimates—rough calculations of how likely Biden is to win the EV from each state, assuming such likelihood is distributed normally:
- For expected margins, I used a mean of estimate-0.8 and a standard error of 7.1
- For WAPA, I used a standard error of 3.0, roughly the margin of error in most quality polls.
- Overall probability Biden wins a state’s EV calculated the same as for predicted final margin
While the means and standard errors are somewhat arbitrary, albeit broadly 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
Six months before election day 2020, and with all of the caveats about what voting will even look like during a pandemic, Biden is clearly in a commanding position to be elected the 46th president of the United States.
- He is projected to win by at least 3.3 points in enough states to get him to 279 EV, or 278 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.
- He has narrower leads in Florida, Ohio and North Carolina, which combine for 62 EV, increasing his total to 340 or 341.
- Arizona’s 11 EV are balanced on a knife’s edge.
- He is favored at least 86% in enough states to get him to 268 or 269 EV
- He would then need to win ONLY ONE of Wisconsin (74.3%), Ohio (72.5%), Florida (69.4%) or North Carolina (67.2%) to win the presidency. Assuming Biden’s chances of winning each state are statistically independent from each other (a lousy assumption), he has about a 99% chance of winning AT LEAST one of these states.
- He has at least a 58% chance in enough states to earn him 351 or 352 EV, at least 81 more than required.
- And if things truly break Biden’s way, he has a 40.5% chance to win the 16 EV in Georgia, a 28.4% to win the 6 EV in Iowa, and a 20.1% chance to win the 38 EV of Texas, upping his total to 411-413 EV, depending on what happens in the 2nd Congressional district of Nebraska, which allocates its EV the same as Maine.
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 335.2, fairly close to the 341 generated by taking the 232 EV won by Hillary Clinton in 2016, adding Michigan and Pennsylvania to get to 268, then adding Wisconsin, Florida, Ohio and North Carolina (and the last EV in Maine).
This lead looks even more robust when you make either of two reasonable assumptions:
All polls are overestimating Biden’s margins by 3.0 points.
In this scenario, Biden’s projected EV drops to 286, still 16 more than required. He would be favored at least 80% to win in enough states to win 239 EV, though he would be favored by at least 64% in three states totaling 30 EV, putting him on the doorstep. He would then have to win one of Wisconsin or Ohio, at 44% each; he would have about a 69% chance to do so.
The point is, even if the polls are consistently off by this much, Biden would still be roughly even money to win the presidency. That said, Biden would still be winning by 3.1 points nationally, demonstrating the current Republican bias in the Electoral College.
All polls are underestimating Biden’s margins by 3.0 points.
In this scenario, Biden’s projected EV are a landslide-level 373.7, more than 100 more than necessary. He would be favored at least 80% to win enough states to earn 341 EV, while being a 77.3% favorite in Arizona and a 69.8% favorite in Georgia, for a total of 368 EV. Adding in the states where Biden would be roughly even money—Iowa and Texas—gets us once again to 412.
This appears to be Biden’s upper limit, as even in this scenario where he is wining nationally by 9.1 points, he is no more than 9% favored to win any additional states.
Now, none of this is to say Biden is guaranteed to be the next president of the United States; it would be monumentally foolish for me to conclude that this far from the election, particularly if Amash earns more than, say, three points in the national popular vote. I am simply noting that all indications point very strongly in that direction, based on the data we have right now.
Until next time…please stay safe and healthy…
 Hawaii, Vermont, Rhode Island, Illinois, Oregon, Louisiana, Nebraska, South Dakota, Arkansas, Idaho, Wyoming
 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