Where Our Model Thinks The Polls Might Be Biased
With less than two weeks until Election Day, Republicans now have a 48 percent chance of taking the Senate, according to FiveThirtyEight’s Deluxe forecast, their highest figure since late July and putting the battle for Senate control officially in toss-up territory. But whether that shift will continue or has peaked is in the eye of the beholder.
Most recently, the GOP hasn’t seen the rapid gains in the forecast that they did a week or two ago, and if you scroll through our latest polls page, you can find a decent number of polls that have good news for Democrats along with quite a few others that look strong for Republicans. Still, the current picture leaves Democrats without much margin to spare. Unless they can pull off an upset in North Carolina, Ohio, or Wisconsin, Democrats will need to win two of the three closest Senate races — Georgia, Pennsylvania and Nevada — in order to maintain their majority, while also holding Arizona and New Hampshire.
In the House, meanwhile, Republicans have an 81 percent chance of taking over the chamber, putting Democrats in a tough — although hardly impossible — position. (In poker terms, it’s as though they have pocket 5’s against pocket aces.) Unlike in the Senate, Democrats will have to do more than just win the majority of toss-up races to control the House. In fact, there are now 219 seats that our model rates as lean Republican, likely Republican or solid Republican, one more than the 218 needed for a majority. So even if Democrats won all the toss-ups, it wouldn’t quite be enough.1 To keep the House, Democrats will need for our model to be systematically underestimating them.
That can happen — forecast and polling errors are correlated. That said, what the polls say and what the FiveThirtyEight forecast says are not quite the same. Most importantly, our Deluxe forecast already assumes the Republicans probably will beat what polls currently show.
Below, you can find a comparison between the FiveThirtyEight polling average for each Senate race where we calculate a polling average, and the projected margin of victory or defeat in the same races according to the Deluxe forecast:
The Deluxe model expects Republicans to beat their polls
FiveThirtyEight polling average and projected margin of victory (or defeat) from FiveThirtyEight Deluxe forecast as of Oct. 27, in 18 Senate races
|State||Polling Average||Deluxe forecast|
In Ohio, for instance, Democrat Tim Ryan trails by only 1.4 percentage points in our polling average, but is projected to lose by 4.5 points, which implies a 3-point polling bias against Republican J.D. Vance.
That’s on the high side for the competitive Senate races, but the Deluxe forecast does expect Republicans to somewhat overperform their polls in most of the other competitive races: by 1 to 2 points in Georgia and North Carolina, and by about 1 point in Wisconsin, Nevada and Pennsylvania. It doesn’t expect much polling bias in Arizona and New Hampshire, by contrast.
In some of the less competitive races, the Deluxe forecasts thinks the bias could be more severe. Check out Iowa, Missouri and Oklahoma, for instance, where Republicans are projected to do much better than their polling — although there are relatively few recent polls of Missouri and Oklahoma. Meanwhile, Democrats are expected to outperform their polls in blue-state races such as Washington, Connecticut, Illinois and New York.
To some degree, this matches the pattern from previous years, although not perfectly so.2 But one fairly good heuristic is simply that Republicans tend to outperform their polls in red states, while Democrats do so in blue states — see the map below for an example of this from 2016.
Why do Republicans do better in the forecast than the polling average? There isn’t any built-in unskewing or assumption of polling bias in the model. Rather, it’s a bunch of little things that tend to add up:
- Our forecast makes additional adjustments to the polls beyond what is shown in our polling averages. In particular, this includes a likely voter adjustment and a timeline adjustment, both of which help Republicans at the margin. The likely voter adjustment helps Republicans because they tend to do better in polls of likely voters than registered ones.3 The timeline adjustment accounts for how the generic ballot has shifted since polls were conducted, and since it’s shifted in a somewhat Republican direction recently, that tends to help the GOP too.
- In some states — Rust Belt swing states such as Pennsylvania, Ohio, Wisconsin, and redder ones like Iowa and Missouri — the “fundamentals” our model considers are more favorable for Republicans than the polls. The most important of these is simply overall state partisanship. Republicans have done fairly well in these states in recent years, so you might expect them to do well again in what is likely to be a fairly good year nationally for the party.
- In addition, the model considers macro-level fundamentals, such as the poor historic performance of the president’s party at the midterms, and President Biden’s mediocre approval ratings. So the model endorses the idea that Democrats have been coming back to gravity. These factors zero out in the model by Election Day, though, so they don’t have too much effect now given how close we are to the election.
- Finally, the expert ratings that the Deluxe model uses are more bullish for Republicans in several key Senate races than polls and other objective indicators.
While none of this is good news for Democrats, it should guard against one potential misinterpretation of our forecast. You might be tempted to take the FiveThirtyEight Deluxe forecast, and then mentally shade it further to Republicans because you worry that the polls could underestimate Republicans again. Heck, I’ve even been tempted to do this myself. But the Deluxe forecast already assumes that Republicans will do better than what current polls show, by around 2 percentage points in the average Congressional race. You may be double-counting, in other words, if you take the Deluxe forecast and then assume additional polling bias on top of it. Some of it is already baked into the model.