We have seen a lot of midterm polls, and to be frank, they look darn good for Democrats: YouGov at Dem +8, PEW at +10, there is even this Quinnipiac poll, at Dem +14. OK, those are all generic ballot polls, and asking if likely voters will vote for “the Democratic or Republican candidate” has only limited predictive power, but other models are similarly optimistic. Nate Silver’s Five-Thirty-Eight gives Democrats above 80% chance to win the election – note that this is a more certain prediction than the above 70% percent chance his blog gave Hillary Clinton to win the 2016 presidential election. So, does this mean progressives can lean back and let victory play out? No! While the Democrats are likely to win the House, the probability that Democrats win the House is much lower than what the pundits are indicating and the probabilities themselves are built on expectations of your participation.

There are two aspects to elections: who will turnout to vote (turnout) and, conditional on turnout, who will voters vote for (preference)? Let us address the latter part first: Conditional on turnout, predicting sentiment, while certainly not easy, is at least easIER. In this highly politicized environment, voters have made up their mind. True, there is still some model error (smart talk for us mis-estimating sentiment conditional on turnout), but we have pretty good reason to believe that in the very least, model errors cancel out in the aggregate, so that our top-lines are close to true. And, room for persuasion, i.e. voters intending to vote Democratic but in fact switching to the Republican camp before the election, are minimal at this point. In fact, most public polling seen at aggregators like Pollster, FiveThirtyEight, and RealClearPoltics, despite showing widely varying generic House polls, are all fairly consistent in the preference of voters, conditional on their demographics.

Now, the former is a much more difficult question to answer. In presidential elections, turnout is more stable. In midterms, the drop-off in turnout, i.e. those turning out in presidential but not in midterm elections, depends on circumstances. We have done some work on understanding what different drop-off scenarios mean, but this research is a what-if scenario using actual turnout data from voter files. Roughly five weeks before the election, we do not have the luxury to rely on actual turnout data.

In short, turnout matters because a vast part of uncertainty of the outcome is now down to the partisan composition of the actual electorate come November. Do Democratic leaning demographic subgroups show up in higher numbers than usual? For example, usually only around 20% of 18-24 year-olds vote in midterms, a highly Democratic group. How high will that number be in 2018? More importantly, turnout is not deterministic, it is instead affected by late-cycle developments, especially in this cycle, and ultimately, whether YOU (and you girlfriend/husband/neighbor) shows up to the polls!

Who is hotter for the Midterms?

Specifically, the key question we face is whether Democratic advantages in enthusiasm about the election translate into partisan differences in turnout on election day. In polling we have done in dozens of the most competitive districts, we have tracked the specific enthusiasm question: “Compared to previous Congressional elections, how enthusiastic are you about voting this year?”. Among self-reported likely or certain voters we register a 12 pp enthusiasm gap for the Democrats (45% Democrats to 33% Republicans that are more enthusiastic than last election). Yet, on other enthusiasm items, the partisan differential does not look quite as strong for Democrats, such as on the item that we would usually consider a generic likely voter screen:
While that chart, stemming from a Washington Post/ABC News Poll, is slightly misleading – the differential here is still a solid 6 percentage points – it is certainly much more modest than the 12 percentage point differential we register. And on other items, the enthusiasm differential also does not favor Democrats quite as heavily as our measure, including:

  • What are the chances of your voting in the election in November 2018?
  • Please rate your level of enthusiasm about voting in the election in November 2018 on a scale from zero to ten.
  • How interested would you say you are in national issues, like elections and legislative debates in Washington?

(Credit to Yair Ghitza here for bringing this to our attention; and one last word on the subject matter: This YouGov poll registers the same percentage of Democrats and Republicans as “certain” to vote in 2018, but this breakdown is problematic in that it counts partisan leaners as Independents, in spite of the outwardly partisan nature and behavior of partisan leaners – not ideal).

In short, based on our down data and most other data that we have seen, we are confident that Democrats are at least somewhat more enthusiastic about these elections than Republicans, although it remains unclear to what extent.

Should we be enthusiastic about the enthusiasm gap?

The obvious question – how much does it matter? Imagine two scenarios: In one, a number of progressive Americans who usually do not vote in midterm elections turn out to vote this cycle, while Republicans who usually do not vote in midterm elections do not turn out and stay home. In the other, a number of progressive Americans who usually do not vote in midterm elections turn out to vote this cycle, but so do a number of Republicans who usually do not vote in midterm elections. It is clear that these two scenarios have differing effects, and the magnitude of that difference cannot be overstated.

To elucidate, we have conducted a little experiment at PredictWise. We have been tracking voting intention of close to 200,000 survey respondents across the country. We project horse race estimates as we always do, using standard PredictWise analytics centered on Mr. P (the affectionate term for our methodology, multilevel regression and post-stratification; as this post focuses on substance over methods, please read here for a detailed description of our polling methods), but we have calibrated our likely voter universe differently to illustrate different scenarios.

Three Scenarios of Turnout

  1. 2014 Scenario (i.e.: not going to happen): This is a very unlikely scenario, but helps putting expectations for 2018 into context. We assume that the composition of the electorate, when it comes to partisan identification, resembles that of 2014, an abysmal year for Democrats. Of all voters in this turnout space identifying with a political party, 48.8 percent of voters identify as Democrats (including leaners), but 51.2 percent of voters identify as Republican, again including leaners. Of course, this would be a disastrous turnout scenario for Democrats and one we think of as extremely unlikely. This scenario sees Democrats loosing the national popular vote by 0.4  percentage points, and sees Democrats effectively winning 4 seats. But, this scenario also shows something else. In 2014, under a similar voter universe, Democrats only took 188 seats and 43.2 percent of the popular vote. It is safe to say that conditional on turnout, voter intention is much, much more favorable to Democrats than it was 2014 (Also, it is useful to remember that the 2014 elections saw the lowest turnout since 1942). 
  2. Baseline Scenario (pessimistic, does not include late-cycle enthusiasm): A likely voter universe derived from turnout models of the biggest Democratic voter file vendor, TargetSmart. This likely voter universe is adjusted for turnout in 2017 and 2018 special elections, but is ultimately built on fundamentals such as previous turnout history and reflects very little of the attitudinal enthusiasm gap we talked about earlier. Of all voters in this likely voters space identifying with a political party, 49.9 percent of voters identify as Democrats (including leaners), and 50.1 percent of voters identify as Republican, again including leaners, according to our estimation. In some ways, as far as turnout is concerned, such a scenario is the most pessimistic realistic outcome for Democrats come November. Assuming that we estimate vote intention conditional on turnout correctly (read more above), this scenario sees Democrats taking the national popular vote by 1.8 percentage points, and sees Democrats effectively winning 9 seats, falling short of the 23 seats they would need to take the House, accounting for currently six vacant seats.
  3. Optimistic Scenario (optimistic, includes big boost from enthusiasm): In this scenario, we assume a substantive correlation of the partisan enthusiasm gap and actual turnout. We calibrate the partisan composition of our likely voter space to this NBC News/Wall Street Journal poll, assuming a radical departure from the composition of the electorate in, for example, 2014. This likely voter universe is not only adjusted for turnout in 2017 and 2018 special elections, but is built more on survey data, taking into account attitudinal differences in enthusiasm to a large extent. Of all voters in this likely voters space identifying with a political party, 55.1 percent of voters identify as Democrats (including leaners), and only 44.9 percent of voters identify as Republican, again including leaners, according to our analytics. This would be an almost historic turnout scenario for Democrats, and we would describe this scenario as the most optimistic outcome possible for Democrats come November. This scenario sees Democrats taking the national popular vote by 9.6 percentage points, and sees Democrats effectively winning 36 seats, taking the House back in a decisive victory. We would call this outcome a wave election.

Take Away

Turnout matters, a lot. Assuming the exact same voting behavior of all voters, three different, realistic turnout scenarios have massive impact on the outcome of the election. Turnout can be the difference between a clear defeat, and a wave election allowing Democrats to take the House back comfortably. And, turnout is not deterministic. Whether November 6 more closely resembles our base pessimistic scenario or our optimistic scenario will depend on canvassing efforts and knocking on doors, and will ultimately hinge on whether progressive-leaning demographic groups that have a higher potential to drop off in midterms will show up this year (we are looking at you, Millennials). Or, in other words, Go Vote! To illustrate the non-deterministic nature, in a few days PredictWise will launch an interactive map that will let our readers explore the effect of different turnout scenarios on the election outcome – both for every district separately and at the national level. We will release the map – developed in collaboration with some of the leading data scientists in that space – on October 6, 2018 – one month before the election!


Of course, this analysis presented here comes with some caveats. First, we assume that we can measure voting conditional on turnout well – true, our models have been very successful in the past, but at the end of the day, this remains an assumption. Second, we use generic ballots to gauge voting intention. In fact, our generic ballot measures closely track candidate-specific, private polling we are conducting this cycle (and the fraction of Americans that as much as know the name of their Congressperson is small), but things could look different under a candidate-specific model. Third, and this is the most complicated point, we adjust our turnout universe uniformly across districts. For example, say our likely voter models attributes a 50 percent likelihood to vote to an individual in a non-swing district in Arkansas, and a 50 percent likelihood to vote to another individual in a competitive district getting tons of media attention, such as California 21. When adjusting our turnout universe to different benchmarks (i.e. the three scenarios we describe above), we effectively tune all individual-level probabilities uniformly, meaning: to the same degree. That most likely is an unrealistic assumption (see for example Nate Cohn’s excellent write-up on accounting for calibrating different turnout universes across districts in the New York Times).

Disclaimer: Politics is not a spectator sport. It is not deterministic. When we say there is 60% likelihood of the Democrats winning the House we are assuming a certain level of commitment by staff, volunteers, and, most importantly, the voters themselves. If Democrats get super motivated: they will win. If Republicans get super motivated: they will win. You are part of the game!