The Electoral College bias has returned with a vengeance

Trump has gained in critical states since the DNC — but that may be reversion to the mean.

Silver Bulletin recently passed 200,000 total (paid + free) subscribers, way more than I ever expected. I’m thrilled that you found us, and if you’re not subscribed yet, I hope you’ll consider joining us for the stretch run of the campaign. 

Just as a quick preview of coming attractions:

• Tomorrow, we’ll have a Model Talk column for paid subscribers on whether prediction markets have a political bias.

• On Sunday or Monday, we’ll have a free column that’s more of a hot take on a certain brand of media criticism that I find really counterproductive.

• And then Tuesday is the first Harris-Trump debate. There may or may not be a preview column — I hate writing previews — but there will definitely be some sort of post-mortem on Tuesday night or Wednesday morning.

---

For today, though, let’s talk about the elephant in the room: the Electoral College is starting to look like a challenge for Kamala Harris. This was a problem for Democrats in 2016, of course, and also in 2020 — when Joe Biden won the popular vote by 4.5 points, but the tipping-point state, Wisconsin, went for Biden by only 0.6 points.

There’s been a lot of talk lately about our model’s convention bounce adjustment, and that’s understandable. Harris is still ever-so-slightly ahead in our polling averages in enough states for her to win 270 electoral votes — but she’s slightly behind in our forecast of those states, mostly because the model assumes that polls conducted just after the DNC are likely to be relatively favorable for Harris.

I think the adjustment is highly defensible, but if you don’t like it, don’t fret: it will begin to work its way out of the model over the course of the next week or two. And if you really hate it, you can always just stick to our “raw” polling averages instead, which aren’t subject to the convention bounce adjustment at all.

The problem for Harris is that Donald Trump has been gaining on her in our polling averages, too — at least in the most important Electoral College states. If, say, Harris had gained 1 point, when the convention bounce adjustment was expecting her to gain 2 points, that would look like more of a rounding error in the model. Instead, though, she’s actually losing ground since the start of the convention in swing state polls.

Take a look at the numbers in the seven most important states, ranked in order of their tipping-point probability. Collectively, there’s an 88 percent chance that one of these states determines the winner, the model calculates. Here are the polling averages as they stood on Aug. 19, the first day of the DNC, as compared to the model run we just published this afternoon:

Harris is ahead by 3.0 points nationally, a tick better than the 2.3-point lead she had coming into the convention. Although I’d note that this is down a bit, too. Her lead peaked at 4.3 points on August 23 in our last model run before Robert F. Kennedy dropped out and we removed him from our model. Still, her national polls are fine; it’s the state polls that are the big issue.

In fact, relative to the start of the DNC, Harris has lost ground in 6 of these 7 states. (Again, there’s no convention bounce adjustment in these polling averages at all; that comes at a later stage.) The notable exception is Georgia, which perhaps makes sense: it had been surprising before that Harris was polling better in North Carolina than in the Peach State. And her numbers are essentially unchanged in Nevada, although with just 6 electoral votes, it only has a 3 percent chance of being the tipping-point state.

But in Pennsylvania, Michigan and Wisconsin, Trump has gained about a point. And the polling average has broken slightly more sharply than that against Harris in North Carolina and Arizona.

What explains this? Let’s work through a couple of theories. First, it’s possible that this reflects some impact from RFK Jr. exiting the race and endorsing Trump, a factor that I think I underplayed initially and frankly that also has probably gotten too little attention in other coverage of the race. RFK is trying — although in several cases, not succeeding — to remove his name from the ballot in swing states, while remaining on the ballot in less competitive ones.

What assumptions do our averages make about Kennedy’s presence or absence on the ballot? Well, basically, we aren’t making any assumptions at all — instead, we’re leaving that up to the pollsters. Our model uses whatever numbers the pollster chooses to publish and averages them together if there’s more than one version (e.g. with and without minor candidates). But even if Kennedy does remain on the ballot in states like Wisconsin, some voters may be interpreting his instructions rather literally, following his endorsement of Trump in states where it makes a difference but not in others.

Another potential factor isn’t my favorite issue to discuss since it makes less difference than people generally assume. But frankly, pollster quality has been a bit higher in national than state polls lately. In principle, our polling averages account for this, with our pollster ratings and our house-effects adjustment. If you think we aren’t going far enough and want to mentally add a couple tenths of a point to Harris’s numbers, I suppose I won’t stop you. But I’d be careful: Harris has gotten some mediocre numbers lately from highly-rated pollsters too, like this one and this one. Other than RealClearPolitics — that’s a subject for another day — most of the more rigorous polling averages show remarkably similar numbers despite having slightly different standards for which polls they include.¹

It’s also possible that Harris’s numbers were overheated in some of these states during the honeymoon phase of her campaign after she replaced Biden on the ticket because of partisan non-response bias. Democrats were extremely enthusiastic after the candidate swap, and although Harris’s enthusiasm numbers are still much better than Biden’s, sentiment has cooled just off a little bit. Obviously, you’d rather have more voter enthusiasm than less, and I can’t emphasize enough how much better a position Democrats are in than they would have been with Biden.

But sometimes enthusiasm translates into unsustainable sugar highs in polls because enthusiastic voters are considerably more likely to respond to surveys. Pennsylvania, Michigan and especially Wisconsin have been subject to significant polling errors in the past two campaigns, and we may be getting a more accurate read on the race now that both Democrats and Republicans are highly and equally enthused about the presidential contest.

Setting all of that aside, I’m going to say something I know our Harris-voting subscribers won’t be happy about: the new numbers look a bit more realistic than the pre-DNC ones given how the country wound up voting in 2016 and 2020.

Take a look at this table. It reflects the partisan lean of each state in 2016 and 2020, and then each state’s projected lean in 2024, calculated in two ways:

• The lean of our current polling average in each state as compared to our national polling average, and,

• The lean of our Nov. 5 projection in each state, as compared to our projection of the national popular vote.

Let me emphasize again, because we’re now firmly in silly season of the campaign where bad-faith actors can dominate the discussion, that these figures reflect the lean of each state relative to the national numbers — and not the raw projections. For instance, our polling averages still have Harris slightly ahead in Wisconsin, but we show it as GOP-leaning relative to the rest of the country.

All seven of these states were GOP-leaning relative to the national popular vote in 2020, and all but Nevada were also redder than the national popular vote in 2016, too. With Biden’s hefty win in the popular vote four years ago, he still won all of them but North Carolina — but Hillary Clinton lost all but Nevada.

So it’s no surprise that both our polling averages and our forecast have these states as GOP-leaning, too. The good news for Harris is that we have the Electoral College bias as being slightly less than in the past two elections. Weighted by each state’s tipping-point probability, it was R +3.7 in 2016 and R +3.5 in 2020. By comparison, our polling averages and our forecast have it at R +2.4 and R +2.5 this time around, respectively.

Still, a two-and-a-half-point Electoral College bias is a lot to overcome in a political environment when most elections are extremely close. And at the risk of piling on bad news for Democrats, I’d note that polls show a considerably smaller GOP lean in Wisconsin than in 2016 and 2020, a possible concern for Harris given the history of polling errors there.

So as you can see from this next chart, we have the Electoral College in toss-up range even if Harris wins the popular vote by between 2 and 3 points. And she’d need to win the popular vote by roughly 4 points to truly “safe” in the Electoral College.

Biden’s polling was terrible, but relatively speaking, his numbers were holding up better with old white voters than among younger voters and among racial and ethnic minorities. That does not mean that Biden was the better candidate: Harris has gained ground on him literally everywhere. But she’s gained relatively more ground among young voters and voters of color, and slightly less among older white voters of the sort that are plentiful in Wisconsin, for instance. This remains an extremely winnable election for Harris, especially with the debate looming next week. But there’s now a 20 percent chance that she wins the popular vote but loses the Electoral College.

1. For what it’s worth, we’re actually more restrictive than 538 at this point. We don’t use polls from ActiVote, Big Data Poll, SoCal Polling and Quantus — but we do use polls from Rasmussen Reports, which they don’t.