Dem 50
image description
Ties 2
GOP 48
image description

Who Aggregates the Aggregators?

Is that headline a reference to an episode of Star Trek: The Next Generation? Why yes, yes it is.

We are getting very close to D-Day (or, maybe, E-Day). Every reader knows that, in the races that will determine control of the Senate, the polls are kinda all over the place. But what about the folks who are aggregating the polls? In theory, if you incorporate a whole bunch of polls into your prediction, you eliminate the fluctuations that might come out of a wonky sample for one poll, or a polling house that uses an... unusual model of the electorate.

So, let's take a look. Our map above has 12 races that are not either strongly Democratic or strongly Republican. Let's toss in Utah, to give us lucky number 13. That will pretty much give us every Senate seat that might conceivably still be "in play." Besides ourselves, we are aware of three sites, namely FiveThirtyEight, RealClearPolitics, and RacetotheWH, that aggregate Senate polls. Here's how the four of us see things:

State Electoral-Vote FiveThirtyEight RealClearPolitics RacetotheWH
Arizona Mark Kelly +1% Kelly +2.3% Kelly +2% Kelly +2.5%
Colorado Michael Bennet +8% Bennet +9.7% Bennet +5.3% Bennet +9.6%
Florida Marco Rubio +6% Rubio +7.1% Rubio +7.5% Rubio +8.1%
Georgia Raphael Warnock +1% Tied Herschel Walker +0.5% Warnock +1.5%
Iowa Chuck Grassley +8% Grassley +7.2% Grassley +7.3% Grassley +10.8%
Nevada Tied Adam Laxalt +0.6% Laxalt +1.9% Laxalt +0.5%
New Hampshire Tied Maggie Hassan +2.9% Dan Bolduc +0.3% Hassan +5.4%
North Carolina Ted Budd +2% Budd +3.1% Budd +5.0% Budd +4.6%
Ohio Tim Ryan +2% J.D. Vance +2.7% Vance +3.3% Vance +4.4%
Pennsylvania John Fetterman +1% Fetterman +0.3% Mehmet Oz +0.3% Fetterman +1.3%
Utah Mike Lee +10% Lee +9.8% Lee +9% Lee +10.6%
Washington Patty Murray +4% Murray +6.1% Murray +3.0% Murray +7.3%
Wisconsin Ron Johnson +3% Johnson +3.8% Johnson +3.2% Johnson +5.8%
Net Result 50 D, 48 R, 2 ties 50 R, 49 D, 1 tie 53 R, 47 D 50 D, 50 R

As it turns out, aggregation has its limits, since—as with the polls themselves—there are a lot of choices being made. For example, we use a fairly simple mathematical model in which we average recent, reliable polls. FiveThirtyEight uses a much more complicated model that weights polls, and considers past electoral history, and requires a Ph.D. in math to fully understand. RacetotheWH uses an approach that's more complex than ours, but less complex than FiveThirtyEight's, weighing polls based on sample size, recency and the pollster's past performance. RealClearPolitics just averages all the polls in their database, which means that, say, outlier polls from March are still influencing their predictions.

Also important is which polls are included. RealClearPolitics is well known for welcoming every Republican-leaning house, no matter how poor their methods are, while (apparently arbitrarily) excluding some polls whose results are favorable to Democrats. As you can see here, they are living up to that reputation this cycle. FiveThirtyEight and RacetotheWH tend to be pretty inclusive, relying on their algorithms to minimize the impact of dubious polling houses. We're actually the most picky of the four; we don't include polls from partisan houses at all (unless one house from each side participates), and we also reject polls from pollsters with poor track records.

In the end, assuming the polls are reasonably on-target (and, thus, that the aggregators are reasonably on-target), it looks like control of the Senate will come down to five seats: Georgia, Nevada, New Hampshire, Ohio and Pennsylvania. Despite the differences between the aggregators, we all agree on eight of the 13 "in play" seats, and in most cases don't have them as being particularly close. If each of those five races proves to be a coin flip, then control of the Senate will come down to that fifth coin flip. However, the results in close races tend to correlate with each other, so 4-1 (or even 5-0) is more likely than 3-2.

Further, there is the small matter that the polls might not be on-target. After all, as a study by Vanderbilt University observed, the 2020 polls had the biggest errors in 40 years. Have the pollsters been able to learn from their mistakes? More on that in the next item. (Z)

This item appeared on Read it Monday through Friday for political and election news, Saturday for answers to reader's questions, and Sunday for letters from readers.                     State polls                     All Senate candidates