Princeton Election Consortium

Innovations in democracy since 2004

Outcome: Biden 306 EV (D+1.2% from toss-up), Senate 50 D (D+1.0%)
Nov 3 polls: Biden 342 EV (D+5.3%), Senate 50-55 D (D+3.9%), House control D+4.6%
Moneyball states: President AZ NE-2 NV, Senate MT ME AK, Legislatures KS TX NC

In late Senate polls, a small signal – or noise?

November 5th, 2018, 8:56pm by Sam Wang

I assume you’ve all been getting out the vote. And donating to one of the organizations in the left sidebar. Maybe you’ve even voted already! OK, now let us take stock of late-breaking developments, which are a little unexpected.

All season I’ve thought that Beto O’Rourke (D-TX) wouldn’t come close to unseating Senator Ted Cruz (R). But the last week of surveys (end dates 10/29 or later, median of n=4 pollsters) show Cruz ahead by only 3.5 +/- 3.1%. That gives me pause.

The 3.1% uncertainty includes 3 points of systematic error by pollsters, based on past midterm elections. That’s a 3.5/3.1 = 1.1 sigma lead, which converts to odds of about 4-1 for Cruz. This is not a slam dunk. I think reports of Beto’s electoral demise are premature. I don’t know who will win, but it could be close.

And the other races? Here’s the last week of surveys (end dates 10/29 or later). Except for Tester, there’s a small but noticeable movement toward Team Blue. It looks like the Kavanaugh bounce has mostly ended.

There are five races within 1 sigma: Indiana, Florida, Arizona, Nevada, and Missouri. All except for Missouri show a slight lead for the Democratic candidate…but with individual win probabilities in the 0.2-0.8 range for either party.

The Meta-Margin is R+4.2%, i.e. overperformance by that much make control a perfect toss-up. Lucas Manning (PEC webmaster) and I will use this data to make a final update to the history tracker.

The systematic (i.e. correlated) error will be known after the election. In the Senate, it usually falls in the direction of Presidential (un)popularity. Democrats could well win all five races, including Missouri (or they could lose all five). If the former happens, that gets Democrats+Independents to 50 seats. In the other direction, an error favoring the President’s party is less likely but would lead to 45 D+I seats.

Of course, Democrats could also fall short. Easy to see that happening, especially in Montana, Missouri, and maybe Indiana. Now we know what Senate races to watch most closely!

And, to state the obvious: if all the close races were to fall the Democrats’ way, the Texas race would become very important indeed.

Tags: 2018 Election · Senate

6 Comments so far ↓

  • TC

    It appears likely that Democrats would win 5 of these 10 races off the top. Beyond that, to get to 51 seats, in the other 5 races they need to win one more than their fair share of coin flips and then pull a supposed upset, for which they have multiple chances to do so. (Even at the outside the increased registration and turnout rates in North Dakota seem encouraging to this end.) It’s only a hunch but Democratic control of the Senate feels like more of a 50/50 proposition than anything else. In my view, a particularly effective Democratic Senate would need to control something like 65 seats, but 51 seat control is better than Republican control.

    • ArcticStones

      51 D+I seats is effective – at least in the sense that it will prevent the confirmation of out-of-the-mainstream judges that lack broad support.

      I am not a betting man, but my bet is that one (possible two) of the following emerge victorious tonight: Beto O’Rourke, Phil Bredesen, Heidi Heitkamp.

  • LondonYoung

    “We can only know the size of the systematic (i.e. correlated) error after the election.”

    Take a rough guess by using previous years ?
    Assume every race has an overall election uncertainty term and a particular term to that race itself and then calibrate?

    I suspect that won’t work well, though, because different races are covered by different mixes of pollsters – each with their own house effects.

    • Sam Wang

      I wrote about this already – in the Senate, it usually goes in the direction of Presidential (un)popularity. The House varies a bit, maybe because the data isn’t as good.

    • LondonYoung

      I saw that you wrote about that effect – and posted the graphic with the most important data – but I am under the impression you haven’t yet tried to model. I know that data is sparse.

      Maybe there isn’t much to be gained beyond just looking at your graph and eyeballing it. Though the usual fear with that method is that I will see what I want.

    • Sam Wang

      I got somewhat granular with this in 2014 for the American Prospect. If one had state-by-state granular approval data, that might help. One problem is how to treat presidential and off-year elections without having to separate them into different categories. Another question is how to consider alternative hypotheses without having too many hypotheses.

Leave a Comment