By this point in the cycle, the conventional wisdom is clear: “polling favors the Democrats, especially in statewide races, because of candidate quality. But fundamentals favor the Republicans“.
This is a notion further exacerbated by the fact that polling badly missed by overestimating Democrats in the previous election cycle, leaving many with a case of buyer’s regret and cementing their (usually unsuccessful) resolve to pay less attention to polling and far more attention to the “fundamentals” of the election cycle.
The problem here is that “fundamentals” is a very nebulous term. To begin with, how much would one consider the generic ballot when compared to Biden’s approval rating in trying to gauge the true environment in November? This is a topic of much debate, and the general school of thought suggests that approval rating is traditionally extremely informative in gauging how the electorate feels, as voters can reliably be backed to vote in line with their sentiments on the president. A vote for the president’s party is a de-facto stamp of approval, by conventional wisdom, while an expression of disapproval is a sign that they may vote for the opposite party.
With this cycle, however, such a thing may not be true due to the larger underlying factors. To begin with, as Nate Cohn of the New York Times and Nate Silver of FiveThirtyEight have both noted, the nation’s voting preferences shift in response to unpopular policy changes, and in this cycle, the biggest change happens to be the reversal of abortion rights across large swaths of the country, thanks to a conservative Supreme Court majority’s ruling in Dobbs v. Jackson. Furthermore, as Cohn notes, former President Donald Trump still commands an extremely high interest level among the electorate, possibly making the midterm even less about Joe Biden and cementing its nature among some voters as a referendum on the two parties.
There’s also another point Nate Silver has already discussed that I believe is worth reiterating — the generic ballot becomes the most important metric to use for informing vote choice as the election nears, because it directly asks the question that people try to estimate from approval polling. In essence, if on election day, Biden is at 44% approval but Democrats are polling at D+2 on the generic ballot, it is actually probably safer to assume that the eventual national vote lines up more closely with the generic ballot than it does with Biden’s approval ratings.
Another open question is whether the environment will degrade in favor of Republicans between now and election day. History certainly suggests that it should; in fact, the FiveThirtyEight forecast currently thinks that the environment will degrade to R+2.6 based on the underlying factors at play. But this hasn’t happened yet, despite the forecast theoretically calling for something similar for months on end now. If anything, the Democratic lead on the generic ballot has widened over the last two weeks, and the FiveThirtyEight modeled estimate has shrunk from R+3.3 to R+2.6 as a result.
Compounding the oddities about this cycle is the fact that state-level polling suggests a massive Democratic overperformance in several of the states that shifted sharply to the right between 2012 and 2020. In Pennsylvania and Ohio, for instance, Democrats are strongly outperforming the margins that the generic ballot would suggest for these states. For example, Ohio is twelve-and-a-half points more Republican than the nation as a whole. A D+1 generic ballot nationally would suggest that an open statewide race should generally result in an R+11 statewide margin. But Democratic candidate Tim Ryan currently leads Republican J.D. Vance by a hair in the polling averages.
Some of this is due to candidate quality. Vance is fairly unpopular in Ohio, with two recent polls showing a 16-point favorability gap between him and Ryan. It is unrealistic to expect that the electorate will always vote the same way regardless of candidates; as Split Ticket has extensively documented, electability is absolutely a real thing and very much matters in elections, and Republicans have done themselves no favors in this race by nominating an untested candidate to go up against a sitting congressman from eastern Ohio.
But then you consider that polls have systematically underestimated Republicans in Ohio for the last four cycles, and you suddenly begin to wonder how much of this is real and how much of this is a mirage induced by bad polling. After all, as Cohn pointed out in his New York Times newsletter, the same warning signs as 2020 are manifesting, as Democratic overperformances are suddenly showing up in exactly the same states that massive polling errors occurred in during the previous presidential cycle.
It is almost certain that there is some very real movement in this race induced by candidate quality; in fact, Split Ticket is aware that internals on both sides suggest Ryan may actually currently be ahead. But it’s not really clear that this is eleven points worth of magnitude, which would have made this the biggest non-incumbent overperformance for any candidate since Jason Kander in Missouri’s 2016 Senate race. Even if this movement is entirely real currently, you might still expect races to eventually line up with “fundamentals”, as partisans begin to converge around their parties as the election nears.
One crude and yet insightful metric to serve as a “regularization” of sorts and get a better picture of where races may end up is to create a blend of fundamentals and polling, similar to an approach suggested by Elliot Li. In essence, take each state’s environment-adjusted partisan lean and add a standard four-point incumbency boost (the value that Split Ticket found incumbency had for the last midterm cycle) where applicable — this would be the state’s “fundamentals” margin, and would reflect what the margin should be with generic candidates running on both sides. Then, combine this with the current FiveThirtyEight polling average, which would ideally encapsulate movement created by candidate quality.
If you assumed there was no statewide error at all in individual Senate polling, you might get a picture like this as the actual predicted map, which might strain credulity when examining some of the state margins suggested here. This is, notably, very different from our actual forecast.
But when you blend the polling with the fundamentals, you begin to get something much more believable, in that the state-level correlations now largely line up with presidential partisanship while also reflecting candidate quality to some degree.
The map here suggests a 51-49 Senate, which is very much within the realm of what you might expect with a D+1 generic ballot. This is not our official forecast (you can see that here), but something like this would not be the least bit surprising for us to see in November based on polling, our current ratings, and state partisanship relative to the nation.
Candidate quality isn’t discarded in this approach at all — we can clearly see candidates like Mehmet Oz in Pennsylvania blowing eminently winnable races through serious underperformances, especially if current trends continue. But it is slightly discounted given exceptionally high levels of polarization, the polling errors that have dominated recent cycles, and the expectation that partisans close rank around their parties as the election nears, and that’s something we are comfortable with saying.
The question here then becomes the following: can you trust generic ballot polling? This is a much tougher question to answer, but Nate Cohn did note in his recent newsletter that the issue that plagued pollsters in 2020, where Democrats (and especially white Democrats) responded at much higher rates than Republicans, is way less prominent now. In 2020, Cohn notes, white Democrats were 20 percent likelier to respond to New York Times polls than white Republicans were. In 2022, they are just five percent likelier of late, which is in line with the fairly accurate 2019 Times polls and not nearly as big of an aberration in terms of response rates.
Does that statistic alone guarantee that national generic ballot polls will be accurate in 2022? Of course not. But it does give us enough comfort to use the national environment as a tool to create crude state-level fundamental baselines. And in a cycle where so much regarding late movement is unknown and the potential for massive state-level polling errors remains at large, a regularizer like this may not be the worst thing in the world to use to get a better picture of how the eventual map may look.
(Ed. Note, 10:29 AM EST: The fundamentals table was updated to include a corrected version that fixed New Hampshire’s fundamentals)
I’m a software engineer and a computer scientist (UC Berkeley class of 2019 BA, class of 2020 MS) who has an interest in machine learning, politics, and electoral data. I’m a partner at Split Ticket, handle our Senate races, and make many kinds of electoral models.
Leave a Reply