My 2018 Early Vote Forecasts
<Editors note: there is no editor here. I've hacked this out with typos and all so you can enjoy!>
Here is my election forecasts based on the early vote. I've tried to be true to the difference-in-difference approach I use to make forecasting. Since Nate Silver has thrown down a gauntlet on the value of early voting analyses, I note that with the exception of Arizona's Senate election, the early vote is generally giving a similar signal as the poll modeling from 538. In that one race, the poll averages suggest a Sinema lead, whereas the early vote favors McSally. I quibble with the poll averages as to who is leading in Nevada and the size of the Florida lead, but these differences are small and not particularly meaningful when the total error of these estimates are accounted for.
Some may come away from this analysis that the early voting doesn't provide any information about an election. I think quite the contrary, naturally. Consider that I've come to the judgement below based solely on an analysis of the early vote. No expensive polls were conducted that probably cost a combined million or so dollars in the states I analyze. I just analyzed data that was produced as a part of election administration. For the most part, the high rent expensive poll averaging method and the low rent early vote method come to the same conclusions. We'll see if we're both right or wrong.
I believe that polling serves as an important source of information, so I am not dismissive of polls or poll averaging. Hopefully this dumb debate about the value of early voting can be laid to rest, we can agree more information is better than less, and learn how to better incorporate both sources of information -- polls and election administration data -- like the innovative directions Catalist and TargetSmart are taking.
There is one other valuable piece of information the early vote has provided leading into this election besides who is leading: that we will have a high turnout election. I heard much at the start of the early voting period that the historic high levels of early voting for a midterm election were just "cannibalization" -- that voters were just casting their ballots earlier. This error was laid to rest when Nevada and Texas surpassed their total 2014 vote in their early vote alone, and other states have had historic numbers that look like a hybrid between a midterm and a presidential election. Perhaps all will fall apart on Election Day and polling places will be ghost towns. The cannibalization crowd has consistently said that early voting would fall off, but that did not happen. The most likely scenario is that voting will continue as it normally does and carry through to Election Day.
Since casual pundits make gross mis-characterizations about how myself and others analyze early voting data, I lead with a long explainer. If you just want the forecasts, scroll to the bottom.
What is "Early Voting"?
I broadly call all forms of voting prior to the election early voting. There are several different flavors of early voting, and the distinctions are blurred across states upon close inspection.
There are a few states, primarily in the Northeast and South, that still have excuse-required absentee voting.
There are states with mail balloting of various guises - no-excuse absentee voting, permanent absentee status where registered voters can request to always vote-by-mail, and all vote-by-mail states and localities, where all registered voters automatically receive a mail ballot.
There are states with in-person early voting where voters can cast ballots in-person at special polling places prior to Election Day. Unlike Election Day polling places, anyone may vote at any of the available early voting locations offered within their community.
States may mix these early voting options. For example, although Colorado is an vote-by-mail state, election officials also offer in-person early voting, and even permit eligible persons to register and vote. North Carolina has excuse-required absentee voting, but offers in-person early voting. Florida has no-excuse absentee voting with a form of semi-permanent absentee ballot status good for two general elections and offers in-person early voting.
I sometimes receive complaints that I refer to the "early vote" in a state that has excuse-required absentee voting. Yes, states may have legal definitions that do not specifically reference in-person early voting. I find these distinctions not particularly meaningful in the big picture. Every state, including the excuse-required states, essentially permits in-person absentee voting at an election office by what is known as a counter ballot, where a voter can visit an election office and request and cast an absentee ballot in-person. Some Virginia localities even open temporary satellite election offices to manage the demand for their excuse-required absentee ballots. Virginia's excuses are fairly expansive compared to some other excuse-required states, so in practice Virginia looks more like a state with in-person early voting than a state with excuse-required absentee voting. A few of the remaining excuse-absentee ballot states also have lenient excuses, with three allowing absentee voting to all persons over a certain age. These states have functionally no-excuse absentee voting for older people. If one want to get lost in the definitions, strictly speaking, Georgia does not have in-person early voting, they have in-person absentee voting, of which more than 1.8 million people took advantage of. Rather than dwell in the minutia of these definitions, I just call all voting that takes place before an election "early".
Early Voting Trends
In-person early voting existed at the country's founding. The election used to extend several days to permit persons living in remote areas to make the journey to a courthouse to cast their vote in-person. This practice was discontinued in 1845 when the federal government mandated that federal elections would take place on a single day - the now familiar first Tuesday following the first Monday in November. The reason for abolishing in-person early voting, was a familiar one to modern voting debates: to curb vote fraud, so that people could not travel to more than one courthouse to cast more than one vote.
People travelling far away could identify a proxy to vote in their place, a practice that was used widely by soldiers during times of war. Proxy voting died with the advent of the secret ballot in the late 1800's. In its place, states adopted the first excuse-required absentee ballot laws. Initially, the only allowable excuse was being in the military, but gradually the excuses were expanded to cover civilians and the list of reasons were expanded beyond simply traveling.
In recent decades, early voting offerings expanded. In 1980 California adopted no-fault absentee voting. Florida, Tennessee, and Texas adopted in-person early voting in 1996. In 2000, Oregon adopted vote-by-mail for all elections. Early voting options continue to proliferate, and states continue to tinker with their laws, mostly to expand offerings, but sometimes to restrict them.
To further delve into the limitations of what the 2018 numbers can tell us, we need to understand the trends in early voting. Early voting as the share of the total electorate has been increasing over time. This is due to more states adopting more expansive early voting options and to a phenomenon that more people over time use an early voting option once a state adopts it.
Since 1972, the US Census Bureau has asked a question about when people vote (there are a few elections in the 1980's where this question wasn't asked). There is an unmistakable upward trend in early voting. There is also an interesting sawtooth pattern, with of more people reporting that they voted before Election Day in a presidential election than a midterm election. I attribute this to the fact the voters usually have more confidence in their presidential vote sooner because there is plenty of information about the election, whereas in a typical midterm election there is less information about candidates and voters need more time to make their choice.
A couple important words about the early vote data. The Census Bureau's data is drawn from a survey, and is different from the numbers reported by election officials which are the actual administrative records of people who voted. The Census Bureau reported 29% of voters self-reported voting early in 2014; election officials reported 33% of voters cast any form of mail ballot or voted in-person early. Why the discrepancy? Polls have survey error, of course, but a deeper explanation is that in many states voters may return a mail ballot on Election Day, and some states even accept mail ballots after the election that are postmarked on Election Day. Although there has been no investigation of this, I believe these people are truthfully telling the Census Bureau they voted on Election Day, whereas election officials truthfully record they voted a mail ballot.
What Does the 2018 Early Voting Data Tell Us?
The preceding discussion should warn causal consumers of early voting data that there is a lot of nuance hidden behind the top line numbers. Early voting is evolving as states' laws change, and voters change when they vote.
Campaigns can have an effect, too, through early voting get-out-the-vote efforts. The word "cannibalization" is often used to describe how campaigns convince their supporters to vote early instead of Election Day. Campaigns can make early voting voter mobilization a part of their campaign strategy. However, if convincing people to vote was as simple as a campaign contact, we'd have much higher turnout in the United States as both parties mobilized their supporters to win. Campaign mobilization efforts undoubtedly affect early voting, but these efforts pale into comparison to the change in laws and change in voters' behavior that they initiate on their own.
What do we know so far as early voting draws to a close and Election Day is upon us?
The strongest signal is high turnout. There were strong signals for unusually high turnout in 2018 midterm election even before the early voting period started, so this should come as no surprise. We had unusually higher turnout for the special elections, unusually high turnout for the primaries, and poll respondents reported unusually high interest in politics and voting for a midterm election. The turnout rate of 37% for those eligible to vote in the 2014 election was the lowest since 1942's 34%. The average midterm turnout rate since 1972 is 39%, so having a higher turnout rate than 2014 was always a likely scenario.
How much higher will 2018 go?
One simple way to estimate the national turnout from the early vote is to do a projection.
So far, election officials have recorded 34.3 million mail and in-person votes in the 2018 election. This number is less than the true current number because there are not current election reports in all offices. There are many outstanding mail ballots in the high volume mail ballot states, which tend to have their early votes come in later than the in-person early vote states, where we get immediate data. Looking at the expected reporting once all the vote is in, there will be somewhere around 40 million mail and in-person votes in the 2018 election.
Election officials reported about 27.5 million early votes in the entire 2014 election. (The 27.5 number is slightly higher than the 27.2 million number on my early vote tracker spreadsheet to account for three states with no 2014 data.)
Calculating a naive estimate of the 2018 total turnout based on a projection from the 2014 early vote yields the following:
In 2014, 33% mail and in-person early votes were cast as a percentage of the total (25.5 million divided by 83.1 million total votes).
Applying the 33% to 40 million yields 121 million votes, for a turnout rate among those eligible to vote of 51.4%.
A turnout rate of 51.4% would not only be higher than the 1966 midterm turnout rate of 48.7%, it would beat the 1914 turnout rate of 50.4%. I would certainly welcome such a high turnout rate for a midterm election, but there are a couple of clear problems here.
First, there are a few states that have adopted more expansive early voting laws than used in 2014 and the use of early voting is trending upwards even in states that haven't changed their laws.
Second, we are seeing what appears to be a hybrid presidential-midterm turnout election. Recall from the discussion above that people vote when they make their choice, and more people tend to vote earlier in a presidential election when there is more information to make their choices. Turnout rates strongly follow patterns of competition in a midterm election, and we are seeing this familiar pattern among the states with competitive U.S. Senate and Governor elections. Something different is going on in this election. Early voting is running at exceptionally high levels for midterm elections in states without these competitive elections driving turnout. The only explanation is that people are voting to express their support or objections to President Trump and his administrations' policies.
Comparing to 2016 yields the following:
In 2016, 42% mail and in-person early votes were cast as a percentage of the total (57.8 million divided by 138.8 million total votes).
Applying 42% to 40 million yields 96 million votes, for a turnout rate among those eligible to vote of 41.1%.
A turnout rate of 41.1% is a lot different than 51.4%, and would equal the 1994 turnout rate.
Given the patterns in the percentage of early voting evident in the Census Bureau data, the actual percentage of early votes in the 2018 general election will lie somewhere between these two extremes.
If I split the difference:
The average of 33% and 42% is 37% (with rounding).
Applying 37% to 40 million yields 107 million votes, for a turnout rate among those eligible to vote of 45% (this number just barely misses getting rounded up).
A turnout rate of 45% seems reasonable, I guess. I am forecasting out-of-sample, which is always a danger for any forecasting. I think there is likely more upside than downside to this estimate since it implicitly assumes a high degree of "cannibalization" whereby polling locations in high volume early voting states are ghost towns on Election Day. The typical pattern of in-person early voting is that it continues to ramp up to the final day, and we observed this pattern again this year. The cannibals weren't running out of voters to eat, so we'll see if the are more bodies on Election Day to consume.
Working with Edison Media Research, we worked through each of the states using a similar methodology that also takes into account comparable past elections, and we arrived at a similar 45%.
What Does Early Voting Tell Us About Who Will Win?
This is what people really want to know, and you've probably scrolled past all the information above to get here.
If you did, let me warn you: we can't tell who will win just by looking at the top-line numbers of who is leading in the early vote. Laws change, voters' behavior changes, and campaign strategies change. We get the strongest signal when there is a past election to compare against. Even, then, there is a good deal of guesswork involved in these forecasts.
I think the strongest signal from the early vote is that we will have an unusually high midterm election turnout rate. I'm seeing some pundits who initially said to ignore early voting being forced to face the fact that the unprecedented high level of early vote for a midterm election is a strong indicator that we will have a high turnout election. There are two states (as I write this) that already exceeded their entire 2014 turnout, both in early and Election Day vote: NV and TX. AZ will likely join them, as well as the all-mail ballot states. There are 28 other states that exceeded their 2014 early vote. Among these, 8 states more than doubled their 2014 vote. These comparisons are complicated by some states adopting new expansive early voting laws, but this cannot be the only explanation. 2018 will be a high turnout midterm election.
What does this mean for election forecasting? For many reasons, election pollsters have to guess who a likely voter will be and will weight their survey to reflect their guess. These are informed guesses, but there is considerable variation among pollsters' guesses. One guess is to assume the electorate will look like the last midterm election, 2014. At this point, this appears to be a poor choice. Some pollsters allow for higher turnout, and I'd put more credence on these polls. You can be a discerning poll consumer by reading through a poll's methodology and look for information about the likely voter model. Unfortunately, not all pollsters reveal their secret sauce, but some do. A new and welcome innovation this election is that some pollsters are publishing their results under different turnout scenarios.
There is more evidence to help us forecast in some specific state election contests. I am of the opinion that the early vote should be another piece of evidence to help election forecasting alongside polls. Unfortunately, the dismissive attitude by some pundits have put me into a position to more strongly defend the technique used by myself and others that suggests I think these are two distinct methods with one being more valuable than the other.
The early vote data that are available in some states is party registration. Some states, primarily those in the South, track race and Hispanic ethnicity. Finally, the crudest available data is geography, which counties or townships the early vote is coming from. Some campaign organizations - such as TargetSmart and Catalist - have posted their partisan modeled data of the early vote, too. Although these organizations are not transparent as to how they enhance and clean their data and their estimation algorithms, more information to consume is always better than less, and perhaps over time we will be able to assess the accuracy of these data.
Simply looking at who is leading in the early vote at a point in time is a poor way to analyze the early vote. Yet, those who wish to dismiss early vote analyses assert this straw man and then argue against it. None of the serious people who analyze early voting use this method.
The best way to do an early vote analysis is to look at a comparable past election and see what the shape of the electorate looks like at a similar point in time. Let's break this down into two parts:
A past comparable election is an important, particularly when early voting laws change.
Analysis is best if it is done at a similar point in time prior to an election since there are predictable patterns to early voting. Republicans tend to cast mail ballots in states with multiple method of voting (the all-mail states are different), while Democrats tend to vote in-person. Young people and people who tend not to affiliate with a party tend to vote later. An analysis conducted before a state's in-person early voting begins will appear more favorable to Republicans that what the final early voting statistics will tell us.
When these two conditions are met, we can have some confidence in what the early vote is telling us. The method I use is called difference-in-difference. In a state with party registration, it involves examining the margin in party registration a party had in a past comparable election and comparing it to the current margin. If the Democrats had a lead of +4 in a prior election and the Republicans have a +4 lead in the current election, there has been a clear shift in the fortunes of the parties and we might infer the Republican candidates are more favored than compared to the past.
Ultimately, we do not know who early voters have voted for. This is where polling data can be important and complementary since it can help reveal expected partisan cross-over vote and how the voters who do not affiliate with a political party will break.
So with that, let me make some predictions, particularly in the states were there is good party registration data.
In 2014, registered Republicans held a +11.5 point lead, and Ducey (R) won by 11.9 points. In 2016, registered Republicans held a +6.7 lead and Trump won by +3.6 points. As of Friday in 2018, registered Republicans led by +7.9 points, but the lead has shrunk each day and we have not gotten the report from the last day of in-person early voting on Friday.
In the governors race where Ducey is running for reelection, 538 points to a large +13.5 point lead for him. I'm confident he leads this race.
In the Senate race, 538 has a Sinema (D) +1.4 point lead. The early vote points to a McSally lead of perhaps 8 points if we think the election will be like 2014, or 5 points if the election is like 2016. Democrats will need to make up ground in the final report on Monday and will need crossover voters or no party affiliated voters to break towards them for them. This could happen, but I'm inclined to believe McSally has a lead (I'll update this with the Monday report when I can).
Registered Republicans usually have a lead in Colorado's early vote. In 2014, registered Republicans had a +7.3 advantage near the end of the early voting period (not including the Election Day activity in this nearly all-mail ballot state). Incumbent Hickenlooper (D) won by +3.4. In 2016, registered Republicans held a +0.4 lead. Clinton won by +4.8 points. As of Saturday, registered Republicans hold a +0.3 point lead, and I expect Democrats to have a net gain on Monday. This doesn't seem like much of an advantage, if there even be one come Monday. %38 has a Polis (D) with a comfortable +6.3 point lead, and there is nothing in the early vote data that contradicts that.
Florida is very close.
In 2016, registered Democrats held a +1.5 point lead in the early vote and Trump (R) won by +1.2 points. In 2014, registered Republicans held a +2.8 lead in the early vote and Scott (R) won +1.1 points. Complicating these numbers is that Florida mail ballot usage increased following 2014 due to a change in the law that stipulates a mail ballot request is good for two general elections.
In 2018, as of Saturday, registered Republicans held a slim +0.6 lead. Democrats may pull ahead in the final Sunday day of in-person early voting.
538 estimates Nelson (D) leads by +2.7 points and Gillum leads by +4.0 points. Given the polling error, the outcome of the election is still uncertain. If we're going to have a hybrid midterm-presidential turnout election, the early vote suggests the election will be close, perhaps even closer than what the polls suggest.
Registered Democrats hold a +7.9 point lead in the early vote. In 2014, registered Democrats led by +1.9 points. It would seem Hubble (D) is well-positioned to defeat incumbent Reynolds (R), even if incumbents are difficult to defeat. Case in point is that Reynolds replaced popular incumbent Branstad (R), who left the governorship in 2016 for a Trump appointment. Branstad crused to victory with a +21.7 point victory in 2014. He
The venerable Des Moines Register poll has Hubble (D) leading by +2 points. 538 has Hubble leading by a narrow 0.4 points. My judgement is Iowa will also likely be close. Reynolds is not as popular as Branstad, but Iowa is only a moderately large early voting state, with perhaps about 40% of its vote being cast early this year.
Looking a little deeper, there has been a lot of interest in King's 4th Congressional district. A recent Times-Sienna poll has King leading +5. His district is ruby red and the shift towards the Democrats in Iowa is coming primarily out of the 1st and 3rd districts. I will not be surprised if King is reelected.
Nevada is very close.
I'll defer to Jon Ralston, who uses the same methodology I do. He is intimately familiar with Nevada politics and is fighting the good fight for journalistic independence in Nevada. The early vote points to narrow wins of +2 points for Rosen (D) and Sisolak (D). The early vote is a little more bullish for the Democrats in Nevada than the polls, as 538 predicts a narrow +0.4 Heller (R) lead and a +1.5 Sisolak lead. There is not a lot of daylight between the estimates given the uncertain nature of both.
What About Georgia, Tennessee and Texas?
These are high volume early voting states, but there is not a lot of data to go by since these states do not have party registration.
Georgia nearly doubled its 2014 early vote, but the racial composition is nearly identical. Whites were 58.1% of the early voters in 2014 and are 58.1% in 2018. African-Americans were 30.8% of early voters in 2014 and are 30.8% in 2014. Who are the White voters? What about more young people voting (they are)? Hard to know for sure where this election stands just based on race demographics.
Tennessee does not publicly release much data on early voting, other than which county is it cast in. We know that the early vote more than doubled 2014, and I unfortunately do not have the time to do in-depth analysis of where the votes are coming from.
There is not a good past election comparison for Texas, so it is hard to know the significance of the eye popping early vote numbers, where more people have voted early in Texas than in the entirely of the 2014 election - early and Election Day votes. Texas has had abysmal turnout for decades - it usually ranks near the bottom nationwide. The Senate election will be a good test of the theory that Democrats can win (or at least narrow the margin) if they can get their voters to participate. A Texas battleground state in a presidential election would dramatically change the Electoral College math.