Category Archives: Uncategorized

Accelerating Election Science

It’s really exciting to write that our report to the National Science Foundation’s Convergence Accelerator (NSF-CA) program, “Election Science: A Proposed NSF Convergence Accelerator” is now available. In this report, we argue that there are fourteen research challenges present critical opportunities for NSF-CA funding.

This report was produced with support from the NSF-CA, and my co-authors (or in NSF lingo, co-Principal Investigators, or “PIs”) on the project are Moon Duchin (Tufts University), Gretchen Macht (University of Rhode Island) and Charles Stewart III (MIT). It was a fun and productive experience working with Moon, Gretchen and Charles.

The project started with a number of brainstorming meetings among the co-PIs, and we thought long and hard about what we each identified as the important research opportunities in election science, challenges that were would require collaborations between technology providers, election officials, and academics — and challenges that if appropriately funded, might be solved within a few years.

We started with a first virtual workshop, which had an innovative format. The co-PIs invited a number of election officials from across the United States, representing different types of election jurisdictions, to sit down with one of us and discuss the pain points that they have experienced in recent election cycles, especially during the 2020 election. We also discussed with them the research priorities that the co-PIs had identified and got their input and feedback on those. We recorded these conversations, which was a unique experience, as it’s rare for academics like us to have the chance to sit with some of the most innovative election officials in the world and have an extended conversation about their experiences and what they think academics can do to help improve American election administration and technology.

These recorded conversations, then, were made available to the participants in four virtual workshops held this spring. Workshop participants represented the diversity of experiences and opinions about election administration and election science: academics (both experts in election science and those new to the area), election officials, technology providers, government officials, and stakeholders.

These four workshops spanned four days and sixteen hours, involving intensive large group, small group, and even one-on-one conversations, as we tried to narrow down the set of research challenges and find areas where there was consensus among the workshop participants about the importance and viability of each potential research challenge. In the end, the co-PIs were able to narrow down the list to fourteen research challenges, each of which is both “convergent” (in the sense that they require active collaboration between academics, and both the private and public sectors) and “ripe for acceleration” (in the sense that with appropriate funding solutions could be build in a short period of time.

The fourteen challenges are:

  1. Evaluating Tools for Election Administration
  2. Ensuring Usability within the Voting Experience
  3. Improving Access to Voting
  4. Communicating Effectively with the Electorate
  5. Detecting Anomalies in Election Management Systems
  6. Sharing Election Results for Research, Dissemination, and Anomaly Detection
  7. Visualizing Election Data
  8. Enhancing Voter Identity Verification
  9. Securing Electronic Ballot Delivery & Return
  10. Implementing End-to-End Verifiability
  11. Improving Cybersecurity for Election Administration
  12. Managing Election Geography
  13. Promoting Sustainable and Scalable Sharing of Election Technology
  14. Developing Next-Generation Voting Technologies

I encourage you to take a closer look at the report to learn more about why we identified each as an important research challenge.

So what happens next? Our report is now being reviewed by the NSF-CA, along with the reports from a number of other NSF-CA initiatives that they supported in this year’s cycle of studies. We hope that the NSF-CA decides that supporting these research challenges in the near future is an important priority, and if they do, that our initiative will move to the next stage of development at the NSF-CA.

I want to again thank the many busy people who took time this spring to work with us to help develop this agenda for accelerating research on election science. We learned a great deal from our workshop participants, and while we were forced to meet virtually because of the pandemic, we put the technology to good used and were able to meet new and get to know people passionate about advancing election science. Finally, I want to thank my co-PIs for a great opportunity to work with them, I learned a great deal about aspects of election administration and technology because of our collaboration.

Why Do “Blue Shifts” Happen in American Elections?

Yimeng Li, Michelle Hyun, and I just posted a revised version of our paper, “Why Do Election Results Change After Election Day? The “Blue Shift” in California Elections”. In the paper, we focus our attention on data from Orange County, CA, part of our ongoing research on election integrity in California. Orange County is a very good case for research, as we have a wealth of amazing data from OC, we’ve learned a great deal about how elections are run in OC, and of course, because the Orange County Registrar of Voters, Neal Kelley, has been wonderful about working with our research group.

Research like this would not be possible without strong partnerships with election officials like Neal Kelley. They are also difficult without a strong research group, and in this case, I’ve enjoyed working on this project with Yimeng and Michelle. Yimeng Li is a graduate student at Caltech, who has worked on a number of election science research projects. Michelle is an undergraduate at Caltech, who started working with us on this project as part of her Summer Undergraduate Research Fellowship (SURF) in 2019. As I’ve said many times in many forums, one of the truly remarkable aspects of being a prof at Caltech is working with our exceptional students!

The abstract of our paper summarizes our work and results well:

The counting of votes in contemporary American elections is usually not completed on Election Night. There has been an increasing tendency for vote shares to shift toward Democratic candidates after Election Day in general elections. In this paper, we study this phenomenon using granular data from Orange County, California. Leveraging snapshots of precinct-level election returns and precinct-level demographic and political composition, we conduct the first full-fledged analysis of the potential drivers of vote share shifts. Utilizing snapshots of individual-level administrative records, we provide the first analysis of the characteristics of voters whose ballots were tallied later versus earlier. Far from being anomalous, the vote share shifts are consistent with underlying precinct voter compositions and the order of individual ballot processing. We find the same underlying drivers in North Carolina and Colorado and discuss the implications of the evolving election administration practices across states for public concerns about election integrity.

Of course, in the paper we also present data from other states to help document the general nature of post-election shifts in vote tabulations in the U.S. They are common, and should not be surprising, because election officials are not processing ballots in random order. Rather, they process ballots in batches as they enter into the tabulation and canvass process, which in a jurisdiction like Orange County means that large batches of by-mail ballots may be processed in groups, while ballots from drop boxes and vote centers are likely to get processed in other batches. Provisional and last-minute ballots then may get processed in their own batches as well. As different types of voters are likely to be casting each type of ballot (by-mail early or late, voting in person, drop box voters, and provisional voters), that implies that we ought to see post-election tabulation results shift from party-to-party, especially in very close elections.

We’ve got more research planned on this topic, and of course, we continue to do lots of research with all of the important data we have collected from Orange County.

Election Day Update From California

It’s a beautiful Election Day, at least here in Southern California!

A few quick morning updates.

  • The California Secretary of State’s latest updates (as of 11/1/2020) notes that 11,161,493 vote-by-mail ballots have been returned (with 22,388,716 issued). They are reporting (as of 11/1/2020) 661,265 in-person votes cast. That’s a total of 11,822,758 ballots returned or cast. Note that the final totals on the number of votes cast in the 2016 presidential election statewide was 14,610,509. It’s quite likely that by the time all of the ballots are in that we’ll exceed the total number of ballots cast in the 2016 presidential election.
  • San Bernardino County (one of the Southern California counties we are tracking) is reporting 463,351 ballots returned as of 11/1/2020. At a similar point in the 2016 presidential election cycle (San Bernardino County reported 221,470 ballots returned). Note that the 2020 mail ballot return numbers in San Bernardino County are currently running about twice what they were in the 2016 presidential election.
  • Orange County (another of our Southern California counties we are monitoring) is now reporting 1,208,924 vote-by-mail ballots returned (of 2,024,785 issued). As of 11/02/2020 OC is reporting a total of 182,841 in-person votes cast, which if we add those to the vote-by-mail ballots returned is 1,391,765 ballots returned or cast in OC.
  • Los Angeles County (yes, we are monitoring LAC as well!) last night reported 146,558 in-person votes cast, and 2,651,717 vote-by-mail ballots returned. From the data we are seeing, turnout this morning at voting centers in LA County is strong, and no doubt, there are still ballots being returned by mail, at drop boxes, and at voting centers throughout the county.

You can see our monitoring reports at Monitoring the Election.

So what are we seeing? Obviously there’s a lot of interest in this election, and so far, California’s voters have responded by returning their vote-by-mail ballots, or by voting in-person. More later today as we get additional data, and after visiting a number of in-person voting centers today.

What’s Michigan to Expect with Early Voting?

Charles Stewart III

The Healthy Elections Project has been running a series of surveys in a half dozen battleground states, asking how voters intend to cast their ballots.  I have previously published thoughts on Arizona, Florida, North Carolina, Pennsylvania, and Wisconsin, with a bonus posting about Georgia (where we are not surveying, ourselves.)  This brief note focuses on Michigan.

First, about the intentions themselves.  The table below shows that there was very little movement in reported intentions across the four waves of the surveys.  An ANOVA test fails to reject the null hypothesis that the four waves are drawn from independent samples, so I proceed using results from the pooled survey, discarding the “don’t knows.”

Vote mode intention among likely voters, including don’t knows
Date

Election Day

Early Mail Don’t know

N

Sept. 4 – Sept. 11

44.0%

7.2% 45.1% 3.8%

484

Sept. 16 – Sept. 25

40.5%

4.3% 49.4% 5.9%

490

Sept. 30 – Oct. 9

37.2%

4.7% 55.6% 2.5%

484

Oct. 14 – Oct. 21

39.8%

10.5% 48.6% 1.0%

491

Total

40.4%

6.7% 49.7% 3.3%

1,949

 

Vote mode intention among likely voters, excluding don’t knows
Date

Election Day

Early Mail

N

Sept. 4 – Sept. 11

45.7%

7.5% 46.8%

465

Sept. 16 – Sept. 25

43.0%

4.5% 52.5%

462

Sept. 30 – Oct. 9

38.2%

4.8% 57.0%

472

Oct. 14 – Oct. 21

40.2%

10.7% 49.1%

486

Total

41.8%

6.9%

51.4%

1,885

(By the way, the overall 95% confidence interval for all four waves combined is around +/- 2.2 points.  For any one wave, it’s around 4.4 points.)

Second, to convert these expectations to raw numbers, we need an estimate of turnout. In most of the other memos, I used a couple of ad hoc methods.  Since then, I have come across the state-level turnout estimates that are produced as a byproduct of the FiveThirtyEight presidential election forecasting model, and so I’ll use those instead.  For Michigan, the low and high estimates are 4.8 million and 6.2 million, respectively, which is what I’ll use here.  However, turnout in 2016 was 4.9 million, and thus I’m certain that the low estimate is too low.

Third, here is what the combination of numbers above imply for the final distribution of votes in Michigan, by mode:

Vote mode, combining four waves of survey
Assumed turnout

Election Day

Early

Mail

4.8 million

2.712M

125k

1.963M

6.2 million

3.503M

161k

2.536M

Fourth, a final calculation needs to be made, if we want to estimate the number of mail ballots that will be requested, since not all mail ballots are returned.  For the states that make their absentee ballot files available, it appears that in 2016, around 85% of requested mail ballots were returned.  If that return rate holds for Michigan, then it should expect between 2.3 million and 3.0 million requests.

There is one data issue that needs to be brought up here.  Michigan does not have a full-bore early voting program.  Rather, it allows in-person absentee voting, but it does not keep track of this separately.  For that reason, and because so few survey respondents indicated they planned to vote in-person before Election Day, I will proceed by considering only mail balloting.

The other data issue is that Michigan, unlike most of the other battleground states, does not provide an easily accessible absentee file to the public, nor even a daily report about these statistics.  For that reason, I’m relying here on Michael McDonald’s irreplaceable reports on his U.S. Elections Project website for these statistics.

As of yesterday, the U.S. Elections Project reports that 3,109,105 ballots had been requested and 2,255,280 mail ballots had been returned.  The deadline for requesting mail ballots is this Friday, although the state (really, the world) has been encouraging voters to request their mail ballots as soon as possible.  Therefore, it seems highly unlikely that anything more than 3.2 million ballots will be requested.

On the returns side, the past week has seen about 66,000 ballots returned per day.  This should pick up in the final days leading up to Election Day.  Nonetheless, at this pace, and with the deadline for receipt on Election Day.  That would yield another 462,000 ballots, or 2.7 million, total.

The following graph summarizes the important calculations.

As for Election Day, this is where the turnout estimate really matters.  If we believe the low-end turnout projections of 4.8 million and that 2.7 million mail ballots returned, that leaves 2.1 million to vote on Election Day.  If the high-end projection of 6.2 is correct, then Michigan would be at 3.5 million on Election Day. If I had to choose, I’d go with the higher estimate.

What’s Pennsylvania to Expect with Early Voting?

Charles Stewart III

 

The Healthy Elections Project has been running a series of surveys in a half dozen battleground states, asking how voters intend to cast their ballots.  I have previously published thoughts on Arizona, Florida, North Carolina and Wisconsin, with a bonus posting about Georgia (where we are not surveying, ourselves.)  This brief note focuses on Pennsylvania

First, as to the intentions themselves.  The table below shows that there was very little movement in reported intentions across the four waves of the surveys.  An ANOVA test fails to reject the null hypothesis that the four waves are drawn from separate samples, so I proceed using results from the pooled survey, discarding the “don’t knows.”

Vote mode intention among likely voters, including don’t knows

Date

Election Day

Early

Mail

Don’t know

N

Sept. 4 – Sept. 11

53.5%

2.6%

39.1%

4.8%

489

Sept. 16 – Sept. 25

56.2%

2.2%

39.3%

2.2%

484

Sept. 30 – Oct. 9

56.8%

1.9%

38.2%

3.1%

492

Oct. 14 – Oct. 21

51.9%

3.5%

41.7%

3.0%

491

Total

54.6%

2.6%

39.6%

3.3%

1,957

 

Vote mode intention among likely voters, excluding don’t knows

Date

Election Day

Early

Mail

N

Sept. 4 – Sept. 11

56.2%

2.8%

41.0%

465

Sept. 16 – Sept. 25

57.5%

2.3%

40.2%

473

Sept. 30 – Oct. 9

58.6%

2.0%

39.4%

477

Oct. 14 – Oct. 21

53.5%

3.6%

43.0%

476

Total

56.5%

2.6%

40.9%

1,892

(By the way, the overall 95% confidence interval for all four waves combined is around +/- 2.2 points.  For any one wave, it’s around 4.4 points.)

Second, to convert these expectations to raw numbers, we need an estimate of turnout. In most of the other memos, I used a couple of ad hoc methods.  Since then, I have come across the state-level turnout estimates that are produced as a byproduct of the FiveThirtyEight presidential election forecasting model, and so I’ll use those instead.  For Pennsylvania, the low and high estimates are 6.0 million and 7.7 million, respectively, which is what I’ll use here.  However, turnout in 2016 was 6.2 million, and thus I’m certain that the low estimate is too low.

Third, here is what the combination of numbers above imply for the final distribution of votes in Pennsylvania, by mode:

 

Vote mode, combining four waves of survey

Assumed turnout

Election Day

Early

Mail

6.0 million

3.390M

156k

2.454M

7.7 million

4.351M

200k

3.149M

Fourth, a final calculation needs to be made, if we want to estimate the number of mail ballots that will be requested, since not all mail ballots are returned.  For the states that make their absentee ballot files available, it appears that in 2016, around 85% of requested mail ballots were returned.  If that return rate holds for Pennsylvania, then it should expect between 2.9 million and 3.7 million requested ballots.

There is one data issue that needs to be brought up here.  Pennsylvania does not have a full-bore early voting program.  Rather, it allows in-person mail voting, which some cities, notably Philadelphia, have promoted.  The Pennsylvania voter file does not have a separate code for in-person absentee ballots, and the conventional way to figure out who was voting in person yielded an implausible number of early in-person voters.  (The standard way is to select records in which the request for the ballot, its issuance, and return all share the same date.)  Therefore, I am assuming that the absentee ballot file does not record in-person absentee voters, although I could be wrong about that.

As of yesterday, 3,058,367 ballots had been requested and 1,850,393 returned a mail ballot.  Today was the deadline for requesting mail ballots, which means that the number of mail ballots requested will likely hit the low end of the estimates, at around 3.1 million.  For the past week, Pennsylvania has been averaging 95,000 returned ballots each day.  Leaving aside the ballots that will arrive after Election Day, if the state keeps receiving ballots at last week’s page, it will eventually have 2.5 million mail ballots, which is also on the low side of the estimates.

The following graph summarizes the important calculations.

As for Election Day, this is where the turnout estimate really matters.  If we think that 2.5 million mail ballots (and a small number of early votes) will come in before Election Day, then if turnout is at the low end of 6.0 million, that leaves 3.5 million to vote on Election Day.  If turnout is at the upper range of 7.7 million, then Election Day turnout would have to be 5.2 million.  This is a considerable range.  My own hunch is that turnout will hit the upper range, so Pennsylvania’s election officials would be well served by planning for this larger number. Whether this is necessary, or an over-reaction, will only be known on Election Day.

What’s Arizona to Expect with Early Voting?

Charles Stewart III

One week to go.

The Healthy Elections Project has been running a series of surveys in a half dozen battleground states, asking how voters intend to cast their ballots.  I have previously published thoughts on Florida, North Carolina and Wisconsin, with a bonus posting about Georgia (where we are not surveying, ourselves.)  This brief note focuses on Arizona.

First, as to the intentions themselves.  The table below shows that there was very little movement in reported intentions across the first three waves of the surveys, but in the last one, there is a move away from Election Day voting, in favor of voting by mail.  For that reason, I’m going to focus this analysis just on the last wave of the survey.

Vote mode intention among likely voters, including don’t knows

Date Election Day Early Mail Don’t know N
Sept. 4 – Sept. 11 24.3% 7.8% 66.1% 1.8% 493
Sept. 16 – Sept. 25 28.7% 10.6% 57.9% 2.9% 477
Sept. 30 – Oct. 9 27.1% 11.4% 58.1% 3.4% 496
Oct. 14 – Oct. 21 16.6% 12.6% 68.8% 2.0% 495

Total

24.1%

10.6%

62.8%

2.5%

1,962

 

Vote mode intention among likely voters, excluding don’t knows

Date

Election Day

Early

Mail

N

Sept. 4 – Sept. 11

24.7%

8.0%

67.3%

484

Sept. 16 – Sept. 25

29.5%

10.9%

59.6%

463

Sept. 30 – Oct. 9

28.1%

11.8%

60.1%

479

Oct. 14 – Oct. 21

16.9%

12.8%

70.3%

485

Total

24.8%

10.9%

64.4%

1,911

(By the way, the overall 95% confidence interval for all four waves combined is around +/- 2.2 points.  For any one wave, it’s around 4.4 points.)

Second, to convert these expectations to raw numbers, we need an estimate of turnout. In previous memos, I’ve done this by adding 10% to 2016 turnout and then utilizing a simple regression model.  Since then, I have come across the state-level turnout estimates that are produced as a byproduct of the FiveThirtyEight presidential election forecasting model, and so I’ll use those instead.  For Arizona, the low and high estimates are 2.7 and 3.7 million, respectively, which is what I’ll use here.  (For the record, a 10% increase would get you 2.7 million and the regression technique would get you 3.5 million.)

Third, here is what the combination of numbers above imply for the final distribution of votes in Arizona, by mode:

  Vote mode, combining four waves of survey
Assumed turnout Election Day Early Mail
2.7 million 456k 345k 1.898M
3.7 million 625k 474k 2.601M

Fourth, a final calculation needs to be made, if we want to estimate the number of mail ballots that will be requested, since not all mail ballots are returned.  For the states that make their absentee ballot files available, it appears that in 2016, around 85% of requested mail ballots were returned.  If that return rate holds for Arizona, then it should expect between 2.2 million and 3.1 million requested ballots. Of course, with the large size of the Arizona permanent absentee list, I should put the word “requested” in quotes (which I just did), because most of these requests didn’t even need to be made for this election.

I’m relying on data being collected and reported on a private website maintained by Data Orbital.  The most recent report from Arizona, which is updated to yesterday (October 26) states that 3,343,186 ballots have been requested and 1,816,615 returned.

One thing to keep in mind is that according to Michael McDonald, Arizona combines early in-person and mail voting statistics, so we can’t distinguish the two.  For that reason, we need to combine both early in-person and mail balloting figures to come up with the comparable statistics that are reported by Arizona.  On the request side, we need to add expected in-person early votes to estimates mail requests, which get us an equivalent of 2.5 million to 3.6 million estimated requests.  We then need to add together the in-person early and mail ballot numbers, to get between 2.2 million and 3.1 million estimated early votes.

Over the past week, Arizona has been averaging 118,000 returned ballots per day.  Applied to the eight remaining days before Election Day, that works out to 944,000 additional ballots to be returned.  If we add this to the 1.8 million already returned, we get an estimate of 2.7 million early votes eventually to be returned, which is within the estimated range from previous paragraph.

These early voting estimates can be summarized in the following graph:

Note that the 2.7 million early votes that I estimate will be returned, if Arizona keeps on its current track, is 100% of the lower estimate for total turnout.  Obviously, Arizona will see some degree of turnout on Election Day, but how much?  The midpoint of the two turnout estimates is 3.2 million.  Taking that as the turnout point estimate, then we get an estimate of 500,000 to turn out on Election Day.  This is  only slightly below the half-way point between the high and low estimates of Election Day voting first calculated above, which suggests that an Election Day turnout level in the range of 500,000 to 600,000 on Election Day would not be out of the question.

What’s Florida to Expect with Early Voting?

Charles Stewart III

It’s just over a week before the election.  The Healthy Elections Project has been running a series of surveys in half a dozen battleground states, asking the intention of registered voters about how they intend to cast their ballots.  I have previously published thoughts on North Carolina and Wisconsin, with a bonus posting about Georgia (where we are not surveying, ourselves, but have survey evidence from elsewhere.)  This brief note focuses on Florida.

To summarize what follow, if we make reasonable assumptions about turnout and take the survey results as a starting point, the actual pace of mail and early voting exceeds what we would expect, which also means that estimated votes on Election Day seem way too small.  To make the math work, we either need to assume that turnout will be around 11.5 million (a 20% increase over 2016, which would be astonishing), that this coming week will see a throttling back on the pace or early and mail balloting, or that survey respondents have been significantly under-reporting their intention to vote on Election Day.  Alas, we won’t know which explanation holds until Election Day, by which time it will be too late to do anything about it.

First, as to the intentions themselves.  The table below shows that there was very little movement in reported intentions across the first four waves of the surveys, but in the last one, there was a clear decline in reported likelihood of voting on Election Day, associated with an increased tendency to prefer voting by mail.  Although an ANOVA test fails to reject the null hypothesis that the percentages from the four waves come from different underlying distributions, the apparent break in mid-October suggests that it might be useful to consider the last set of results separately.

Vote mode intention among likely voters, including don’t knows

Date

Election Day

Early

Mail

Don’t know

N

Sept. 4 – Sept. 11

29.2%

24.5%

41.8%

4.5%

493

Sept. 16 – Sept. 25

29.5%

30.3%

38.2%

2.1%

486

Sept. 30 – Oct. 9

25.7%

32.2%

39.7%

2.4%

489

Oct. 14 – Oct. 21

21.3%

31.2%

46.6%

0.9%

492

Total

26.4%

29.5%

41.6%

2.5%

1,960

 

Vote mode intention among likely voters, excluding don’t knows

Date

Election Day

Early

Mail

N

Sept. 4 – Sept. 11

30.6%

25.6%

43.8%

471

Sept. 16 – Sept. 25

30.1%

30.9%

39.0%

476

Sept. 30 – Oct. 9

26.3%

33.0%

40.7%

477

Oct. 14 – Oct. 21

21.5%

31.5%

47.0%

487

Total

27.1%

30.3%

42.6%

1,911

(By the way, the overall 95% confidence interval for all four waves combined is around +/- 2.2 points.  For any one wave, it’s around 4.4 points.)

Second, to convert these expectations to raw numbers, we need an estimate of turnout.  I’ve done this in two ways, as I have for all the states in the previous memos.  The first is just to assume that turnout is 10% above the 2016 turnout level of 9,580,489.  This would place 2020 turnout at 10.5 million. The second uses a very simple linear regression to predict the log of turnout in terms of the log of registered voters in the most recent six presidential elections, and then extrapolating based on the coefficients derived from that regression.  This gives us an estimated turnout level of 10.8 million. (For those paying close attention to the various posts on this subject, the difference between the two estimates are the tiniest yet.)

Third, here is what the combination of numbers above imply for the final distribution of votes in Florida, by mode:

 

Vote mode, combining four waves of survey

 

Vote mode, using last wave of survey

Assumed turnout

Election Day

Early

Mail

 

Election Day

Early

Mail

10.5 million

2.846M

3.182M

4.473M

 

2.258M

3.308M

4.935M

10.8 million

2.927M

3.272M

4.601M

 

2.322M

3.402M

5.076M

Fourth, a final calculation needs to be made, if we want to estimate the number of mail ballots that will be requested, since not all mail ballots are returned.  In the 2016 election, 86% of the requested mail ballots were returned.  If this holds for 2020, then Florida should expect between 5.2 million and 5.9 million ballots to be requested.

The most recent report from Florida, which is updated to yesterday (October 25) states that 5,988,213 mail ballots were either unreturned or had been counted, which I take to be the estimated number of mail ballots that have been requested.  Because the deadline to request mail ballots has now passed, it seems that the number of mail ballots requested will be just above the upper level estimated here.

Also, as of this writing, 3,805,775 mail ballots have been returned. Over the past week, Florida has been receiving an average of 173,000 mail ballots each day.  If this pace continues—and we actually can expect for the pace to quicken—Florida would have 5.1 million ballots in hand by election day, which is also just above the upper limit of the estimates here.

This brings us to early voting.  Over two million Floridians cast ballots in the first week of early voting, with seven more days to go.  Even at this pace, Florida will see more than four million early votes—and we know that the early voting pace will pick up, too.  Thus, early voting is likely to exceed even the upper bounds of these estimates by around 600,000 voters.

These early voting estimates can be summarized in the following graph:

As noted, Florida is already right above the high end of these estimates for mail ballots and is on the way to do the same for early voting.  Where does this leave Election Day?

Let us assume a turnout level of 10.8 million.  If we subtract 5.1 million mail ballots and 4.0 million early votes,  which is the path the state is currently on, that leaves a paltry 1.7 million ballots to be cast on Election Day.  That’s an astonishingly small number, by any standard.  Of course, it all depends on the turnout model and the accuracy of the extrapolations I’ve been suggesting.  Here are some additional thoughts.

  • A 10% increase in turnout over 2016 is already pretty bit, but what if it’s even bigger, at 20%? Then, we’d expect 11.5 million to vote, leaving 2.4 million to vote on Election Day.  That would get us to the lower range of the estimates, but only because we’ve assumed an astounding turnout level.
  • What if the early- and mail voting are more front loaded than in past years? In that case, the pace of mail and early voting might throttle back in the next week, giving more space for Election Day voting.

These two thoughts take off from the current level of mail and early voting and question whether certain assumptions are right.  It’s also the case that we could go back to the survey responses and ask why the proportions voting using the four modes is so off, especially the first three.  That’s the subject of another post.  A favorite hypothesis of mine right now is that both Democrats and Republicans have been responding to these polls reflecting on the clear preferences of their party’s political leaders.  If so, then it does raise important questions about using survey research to anticipate voter behavior in the realm of election administration.

 

 

What’s Georgia to Expect with Early Voting?

Charles Stewart III

I’ve been working on estimates about the number of mail ballots and early ballots we can expect in the battleground states where the Healthy Elections Project has been surveying about voter intentions.  I’ve published estimates for North Carolina and Wisconsin, and should have estimates for Michigan, Pennsylvania, Arizona, Florida, and Ohio shortly.

In hindsight, there’s at least one battleground state where I wish I had been surveying in, and that’s Georgia.  However, there is some other data available we can use, from the Voter Study Group, which can help provide some guidance to where things will be going.

Before jumping in, let me give the spoiler alert.  The survey research I have seen that has asked Georgians how they plan to vote–early, by mail, or on Election Day–seems to be under-predicting early voting turnout and over-predicting how many will cast votes by mail.  Whether Election Day ends up being relatively quiet, with voters siphoned off to early and absentee voting, or busy, with historic turnout rates continuing through to November 3, depends on what your model of the turnout level is.

First, as to the intentions themselves.  According to the VSG’s October 23 update, utilizing data from August 27 to October 21, 30% of Georgians plan to vote in person on Election Day, 36% early, 32% by mail, and 2% don’t know.  Following the convention in my other analyses, I omit the “don’t knows,” leaving us with 30.6% on Election Day, 36.7% early, 32.7% by mail,

Second, to convert these expectations to raw numbers, we need an estimate of turnout.  I’ve done this in two ways.  The first is just to assume that turnout is 10% above 2016 (4,165,405), which would place it at 4.6 million.  The second uses a very simple linear regression to predict the log of turnout in terms of the log of registered voters in the most recent six presidential elections, and then extrapolating based on the coefficients derived from that regression.  This gives us an estimated turnout of 6.0 million. 

Yes, I know this is a huge range.  What makes Georgia’s turnout rate tricky to estimate, using the registration rate, is that its adoption of automatic voter registration has caused its registration numbers to balloon since 2016.  The regression estimate assumes that turnout of registered voters who would not have been registered except for AVR is the same as those who would have registered otherwise.  That seems unlikely, but for now, I’m sticking with the high estimate, simply because it’s probably better under these circumstances to over-estimate turnout than under-estimate it.

Third, here is what the combination of numbers above imply for the final distribution of votes in Georgia, by mode:

 

Vote Mode

Assumed turnout

Election Day

Early

Mail

4.6 million

1.408M

1.688M

1.504M

6.0 million

1.836M

2.202M

1.962M

Fourth, a final calculation needs to be made, if we want to estimate the number of mail ballots that will be requested, since not all mail ballots are returned.  The return rate of mail ballots in the 2016 general election was 88%.  If that holds for the general election in Georgia, it should expect between 1.7 and 2.2 million mail ballots to be requested.

As of yesterday (October 25), 1,539,302 mail ballots have been requested.  Last week saw an average of 7,700 new mail requests each day.  If that is the pace of requests in the coming week (the deadline for submitting an absentee request is Friday), Georgia won’t even see 1.6 million requests for mail ballots. In other words, it appears that Georgia will under-shoot even the low-ball estimate.

If requests appear to be coming in below the estimate, then the number of mail ballots requested are likely to, as well.  As of yesterday, 963,616 ballots had been returned.  To reach the low end of the estimated range of absentee ballots, Georgia would need to see about 79,000 ballots returned each business day between now and November 3.  That would be nearly double the pace of past week.  This is not out of the range of the possible, but it would still put Georgia on a track for 1.5 million mail ballots cast, which is right at the low end of estimates, assuming the survey results are correct and Georgia gets turnout of 4.6 million.

The big story is early voting, and here the pace exceeds poll-based expectations.  As of yesterday, over 1,792,000 early votes had been cast, which means that it has already pierced the low-end estimate of 1,688,000.  For the past week, Georgia has been averaging nearly 135,000 early votes a day.  With five more days left for early voting, that pace would add another 675,000 voters, pegging total early voting turnout at around 2,467,000, which is well above the high-end estimate of 2.2 million.

However, there’s every reason to believe that the pace of early voting will accelerate even further next week.  This expectation is based on the pattern of early voting thus far, which is illustrated by the accompanying graph, which compares daily early voting numbers for 2016 and 2020.  (Click on the graph to enbiggen.) Weekday early voting volumes have ranged from 29% to 70% higher than the comparable day in 2016.  Furthermore, early voting picked up the pace in the final week before the election in 2016, crescendoing to a quarter million on the last day.

If we assume that daily early voting turnout will continue running about 40% above last year’s pace, as it did last week, then Georgia is on a path to add an additional 1.4 million early votes by the end of this week.  Combined with the 1,792,000 early votes already in the ballot boxes, we could see around 3.2 million early votes cast in Georgia when it is all over.

All told, it is looking like Georgia is on a path to see 3.7 million advanced votes this cycle, 1.5 million by mail and 3.2 million early in person.  (In contrast, it saw nearly 2.4 million advance votes in 2016, with total turnout at near 4.2 million.) 

The following graph summarizes these estimates.

Where does this leave Election Day itself?

This is where the turnout estimate really matters.  Simply assuming that turnout is 10% above 2016—in other words, slightly above population growth—gives us a turnout estimate of 4.6 million, leaving only 900,000 voters for Election Day.  The regression-based estimate of 6.0 million leaves a much larger number, 2.3 million.  For context, nearly 1.8 million cast Election Day votes in 2016.

Although we do not know what total turnout will be, nor how many early and mail ballots will be cast next week, it already seems obvious that the Voter Study Group estimates are off by quite a lot. Those estimates are consistent with the volume of early votes being only about 10% greater than the volume of mail votes.  But, the trajectories are currently putting early votes at twice the rate of mail votes.  This is no particular dig at the Voter Study Group, whose results are entirely consistent with other polling that has been shared with me in confidence. 

Nor does this appear to be simply a matter of voters intending to vote by mail and then switching to in-person early voting.  Thus far, around 10% of early voters are recorded as having previously requested a mail ballot.  It’s certainly true that if we reallocate 10% of the estimated early voters to the mail ballot group, the early/absentee voting gap closes, but the relative dominance of early voting would still be roughly a factor of two, rather than parity.

This is a puzzle to be resolved by the researchers after the election.  Where does it leave the state, as it plans for the next eight days? First, it seems obvious that the state needs to brace for even greater early voting volume this coming week.  Second, it also seems that the state cannot be complacent about Election Day.  The cruel reality this year is that we won’t know if this is a high turnout election, or a high HIGH turnout election until the polls close on November 3. 

What’s Wisconsin to Expect with Early Voting?

Charles Stewart III

It’s now twelve days to the general election.  The Healthy Elections Project has been running a series of surveys in half a dozen battleground states, asking the intention of registered voters about how they intend to cast their ballots.  This brief note focuses on Wisconsin.

First, as to the intentions themselves.  Not surprisingly, as the following table demonstrates, as Election Day has approached, voters have become more certain of what they intend to do.  Overall, as the uncertain voters have decided what to do, they have tended to gravitate toward voting by mail.  This is a small trend, based on small numbers, so I wouldn’t make too much of it.  If we take the average across all three waves of the survey and exclude those who don’t know how they will vote, we can expect 44% on Election Day, 12% during in-person absentee (early) voting, and 44% by mail.  If we take just the most recent wave, these numbers are 42%, 10%, and 48%, respectively.

Vote mode intention among likely voters, including don’t knows

Date

Election Day

Early

Mail

Don’t know

N

Sept. 4 – Sept. 11

44.7%

11.1%

40.3%

4.0%

496

Sept. 16 – Sept. 25

42.7%

12.6%

41.6%

3.1%

496

Sept. 30 – Oct. 9

40.8%

9.9%

46.2%

3.1%

494

Total

42.7%

11.2%

42.7%

3.4%

1,486

 

Vote mode intention among likely voters, excluding don’t knows

Date

Election Day

Early

Mail

N

Sept. 4 – Sept. 11

46.5%

11.6%

41.9%

476

Sept. 16 – Sept. 25

44.1%

13.1%

42.9%

480

Sept. 30 – Oct. 9

42.1%

10.2%

47.7%

479

Total

44.2%

11.6%

44.2%

1,435

(By the way, the overall 95% confidence interval for all three waves combined is around +/- 2.5 points.  For any one wave, it’s around 4.4 points.)

Second, to convert these expectations to raw numbers, we need an estimate of turnout.  I’ve done this in two ways.  The first is just to assume that turnout is 10% above 2016, which would place it at 3.3 million.  The second uses a very simple linear regression to predict the log of turnout in terms of the log of registered voters in the most recent six presidential elections, and then extrapolating based on the coefficients derived from that regression.  This gives us an estimated turnout of 2.8 million.

Third, here is what the combination of numbers above imply for the final distribution of votes in Wisconsin, by mode:

 

 

Vote mode, combining three waves of survey

 

Vote mode, using most recent wave of survey

Assumed turnout

Election Day

Early

Mail

 

Election Day

Early

Mail

2.8 million

1.238M

0.325M

1.238M

 

1.179M

0.286M

1.336M

3.3 million

1.459M

0.383M

1.459M

 

1.389M

0.337M

1.475M

Fourth, a final calculation needs to be made, if we want to estimate the number of mail ballots that will be requested, since not all mail ballots are returned.  During the summer, it seems that 90% of requested mail ballots were returned.  If that holds for the general election in Wisconsin, it should expect between 1.4 and 1.6 million mail ballots to be requested.

As of this morning (October 21), 1,419,484 mail ballots have been requested.  In recent weeks, requests have been running at around 10,000 per day, so Wisconsin is on target for around 1.5 million mail ballots.

Also, as of this morning, 947,811 mail ballots have been returned.  This means that about 60% of the ballots that will be returned have been returned.  It also suggests that between now and Election Day, Wisconsin should be seeing an average of 60,000 mail ballots arrive each day, which is just a bit above what was received this past Monday.

One final thing:  Today was the first day Wisconsin started separating out in-person absentee voting from is daily absentee report.  In this first day, nearly 80,000 people voted early.  That’s much higher than the number of mail ballots that were returned (30,846).

Here is the graph that summarizes this all.

 

 

 

 

 

North Carolina Is on Track to More than 1.3M Mail Ballots and 2.1M Early Votes

What’s North Carolina to expect?

Charles Stewart III

It’s now two weeks and a day to the general election.  The Healthy Elections Project has been running a series of surveys in half a dozen battleground states, asking the intention of registered voters about how they intend to cast their ballots.  This brief note focuses on North Carolina.

First, as to the intentions themselves.  Not surprisingly, as the following table demonstrates, as Election Day has approached, voters have become more certain of what they intend to do.  Overall, as the uncertain voters have decided what to do, the fraction saying they intend to vote during the early voting (one-stop absentee) period has grown.  This is a small trend, based on small numbers, so I wouldn’t make too much of it.  If we take the average across all three waves of the survey and exclude those who don’t know how they will vote, we can expect 30% on Election Day, 43% during one-stop (early) voting, and 27% by mail.  If we take just the most recent wave, these numbers are 27%, 46%, and 27%, respectively.

Vote mode intention among likely voters, including don’t knows

Date

Election Day

Early

Mail

Don’t know

N

Sept. 4 – Sept. 11

29.8%

39.8%

24.7%

5.7%

483

Sept. 16 – Sept. 25

30.3%

38.7%

26.8%

4.2%

487

Sept. 30 – Oct. 9

26.6%

45.4%

26.3%

1.7%

488

Total

28.9%

41.3%

26.0%

3.9%

1,458

 

Vote mode intention among likely voters, excluding don’t knows

Date

Election Day

Early

Mail

N

Sept. 4 – Sept. 11

31.6%

42.2%

26.2%

455

Sept. 16 – Sept. 25

31.6%

40.4%

28.0%

467

Sept. 30 – Oct. 9

27.1%

46.2%

26.7%

480

Total

30.1%

43.0%

27.0%

1,402

(By the way, the overall 95% confidence interval for all three waves combined is around +/- 2.5 points.  For any one wave, it’s around 4.4 points.)

Second, to convert these expectations to raw numbers, we need an estimate of turnout.  I’ve done this in two ways.  The first is just to assume that turnout is 10% above 2016, which would place it at 5.0 million.  The second uses a very simple linear regression to predict the log of turnout in terms of the log of registered voters in the most recent six presidential elections, and then extrapolating based on the coefficients derived from that regression.  This gives us an estimated turnout of 5.3 million.

Third, here is what the combination of numbers above imply for the final distribution of votes in North Carolina, by mode:

 

Vote mode, combining three waves of survey

 

Vote mode, using most recent wave of survey

Assumed turnout

Election Day

Early

Mail

 

Election Day

Early

Mail

5.0 million

1.500M

2.150M

1.350M

 

1.350M

2.300M

1.350M

5.3 million

1.590M

2.279M

1.431M

 

1.431M

2.438M

1.431M

Fourth, a final calculation needs to be made, if we want to estimate the number of mail ballots that will be requested, since not all mail ballots are returned.  During the summer, it seems that 90% of requested mail ballots were returned.  If that holds for the general election in North Carolina, it should expect between 1.5 and 1.6 million mail ballots to be requested.

As of this morning (October 19), 1,331,050 mail ballots have been requested.  In recent weeks, requests have been running at around 100,00 per week, so North Carolina is on target for the 1.5 – 1.6 million requests estimate.

Also, as of this morning, 622,781 mail ballots have been returned and 920,337 early ballots have been cast.  On the mail ballot side, this means that about 45% of the ballots that will be returned have been returned.  It also suggests that between now and Election Day, North Carolina should be seeing an average of 50,000 mail ballots arrive each day.  Last week’s peak days saw just over 25,000 ballots received, so from this point out, the average day moving forward will be twice as busy as last week’s peaks.

On the early voting side, with 920,337 votes cast already, this leaves between 1.2 and 1.5 million votes go to.  In other words, North Carolina is also a bit over 40% of the way to processing its early voting, as it is its mail votes.

Here is the graph that summarizes this all.