PEC (and polling) performance review, 2022

November 20, 2022 by Sam Wang

Generally, pre-election analytics did great, as long as they were driven by polling and election data. Here’s a rundown of how we did.

In a nutshell, sticking close to data did great, but reporter and pundit priors tended to vitiate the exercise. Which leads to the question: what should you do with data? (Spoiler: don’t watch the horserace, optimize your efforts!)

First, a general note. Everyone’s using the same data, publicly-available polls. The people at FiveThirtyEight curate this quite well. At that point, the question is how to process the data.

They take a bespoke approach of rating pollsters and building a model with hidden parts (corrections for pollsters and so on). Here at PEC we take a median of recent polls, relying on the median to reduce the effect of outliers. The basic idea is that all the hooha about pollster quality, crosstabs, and so on is overkill. The null hypothesis is that we can automate things and, with careful assumptions about uncertainty, come close in performance.

With that, let’s take a look.

Senate overall polling error: Democrats overperformed polls by 2 points.

This graph shows two facts. First, across states, the true margin generally was more favorable to Democrats than polls, by a median of 2.1 points. That is a normal polling error, in fact smaller than 2018 and 2020. Second, the slope of the points is steeper than equality (the red line), which illustrates the tendency of winning candidates to overperform their polls in either direction.

Together, these observations are consistent with a slight tendency for pollsters to underestimate relative enthusiasm by Democrats. The interesting thing is that it occurred not just in blue states (which is known from past patterns) but also in swing states. Perhaps an effect relating to a focus by some on abortion and democracy?

The three “misses,” in which the winner was not the leader, all occurred in states where polls showed a margin of less than 2 points: Nevada, Georgia, and Pennsylvania. That is a perfectly fine outcome. The only reasonable interpretation of a 1-point lead is to take it as an exhortation to work harder.

Senate aggregate: PEC polls-only, 47 to 51 Democratic seats. Outcome: 50 or 51 seats, depending on the Georgia runoff. This is to be expected. Over the last half dozen Senate elections, the outcome has tended for close races to all fall the same way. In this case it means that either 47 or 51 Democratic seats would be a less surprising outcome than 49 seats.

House aggregate: generic polls, R+2%. Outcome: pending, looks like R+2%. So that came in on the button.

Special election-based prediction: suggested D+3%. Outcome: (raspberry sound) That measure has a big uncertainty associated with it. Probably best to use this as a loose prior.

Gerrymandering/redistricting advantage: In pre-election calculations, I estimated that the overall national map gave Republicans a 2-point advantage in terms of the popular vote. In other words, they could win the House with as much as a 2-point popular-vote loss. That is manifestly untrue, considering the closeness of the outcome (221-222 seats for Republicans, 213-214 seats for Democrats). This suggests that Democrats overperformed in key swing districts.

Optimizing your donations and activism: knife-edge races and races affecting democracy itself. In the aggregate, we did quite well. Throughout the season, I suggested 21 close Senate, House, judicial, and state offices (as well as some state parties). Democrats won 10 of those races (5 Senate, 2 governor, 2 secretary of state, and 1 attorney general) and Republicans won 11 races (3 Senate, 7 judicial, and 1 attorney general). An even split is exactly where you want to be for optimizing donations and efforts.

The secretary-of-state and attorney-general wins by Democrats were important for protecting the conduct of the 2024 election. The judicial races are notable: the picture for voting rights just got bleaker in Ohio and North Carolina.

With that, thank you for your readership this year! We operated on a smaller scale this year because of distractions and staff turnover. I appreciate those of you who have stayed with PEC. Post-horserace, I’ll be spending more time on fixes for democracy. See my Substack and the Electoral Innovation Lab (coming soon!).


Leave a Reply

Your email address will not be published. Required fields are marked *