One of the most striking things about following this campaign is the fact that election projections are all over the place. 

I was chatting with a political science professor friend of mine, and he was projecting a 1980 Reagan-style victory for Trump--an assessment I tend to agree with--while the online analysts are showing a dead heat with a slight advantage to Kamala Harris. 

While Nate Silver has been more Trump-bullish, even his latest projection shows a very slight advantage for Harris. 

The answer, of course, comes down to the polls and what you take away from them. Most of the polls show Harris with a very modest lead in the national horserace, although one that doesn't match Biden's in 2020. That, in itself, should make you a bit skittish about believing that these polls are favorable to Harris. If Trump's numbers are better today than in 2020, and Harris is doing worse with labor, blacks, and Hispanics, how do these polls make sense?

One explanation is simple: pollsters use models, and models are by their nature always wrong. They are guesses about what the electorate will look like, and the pollsters use those models to shape the numbers after they collect the data, "correcting" for non-randomness in their poll samples.

This leads to sometimes massive systematic errors. Assuming that pollsters aren't massaging the numbers to benefit one candidate or another--you make a judgment about that--you still have the problem that you need to use some statistical voodoo to come up with a model of the electorate. 

How many Republicans will vote? How many Democrats? What will the "leaners" do? How many "leaners" will there be? 

It's black magic, not science. Some pollsters are better than others, but since every election is different, you can't just use a previous electorate to model this one's. 

Then you have the "nonresponse" bias. Can you get a random sample if the people who answer the polls are not randomly selected? Obviously not, and since there are indications that some people (Democrats, mostly) are more likely to answer a pollster's question than others (Republicans or MAGA-types), are you getting a representative sample in the first place? 

There are a number of variables like this that influence poll results, and every pollster has their own black magic formula to correct for errors. 

Let's look at PA, where Biden won +1. The @nytimes PA sample shows recalled '20 vote as Biden +8. @EmersonPolling PA sample shows Biden +1 in recalled vote. Franklin & Marshall is Biden +2 but projects 19% of '24 electorate will be non-2020 voters. Three very different pools. 2/3

— Brad Todd (@BradOnMessage) September 19, 2024

Then, you get into the models themselves. Some are more straightforward--projecting victories based solely on poll numbers--and many use their own version of black magic to massage the numbers to make projections. 

I share polls because, sure, I look at them compulsively, & what is social media other than a way to alleviate that & share neuroses collectively. But the false sense of certainty, & overanalyzing of crosstabs or shifts of one percentage points, should def be avoided!

— Taniel (@Taniel) September 20, 2024

Now add another confounding factor: there has NEVER been a campaign remotely similar to this one. Candidate Switcheroos, an unpopular VP, suddenly getting tongue baths from the media, two assassination attempts on Donald Trump, the lawfare, the heated rhetoric, and the Roe v. Wade factor. 

Call it the Black Swan election, or the unicorn, or the zebra participating in the Kentucky Derby. Call it whatever you want, but don't call it modelable. It isn't. 

There are a few things we DO know: in the past, the polls missed out on significant support for Donald Trump--twice. That doesn't guarantee they haven't corrected for that, so the modelers who add some "X" factor to Trump's polls are betting that the same errors will pop up again. They are likely correct, but I wouldn't bet a ton of money on it. 

So why do I feel good about Trump's prospects, as does my former colleague Professor X? 

It's hard to find Harris voters on the ground, and plenty of evidence supports that. 

Don Lemon HUMILIATED as Voters ALL BACK TRUMP!!!

pic.twitter.com/26BEA2hYAo

— . (@Nan1_0o) September 19, 2024

When reporters go out to find Harris voters, they have an unusually hard time doing so. Man-on-the-street interviews are rarely that useful, but it's not like many people haven't tried this experiment and had the same experience. 

This might be my favorite clip from the entire election cycle. CBS visited voters at a diner in Washoe County, NV that went for Biden by about 4.5% in 2020. CBS could only find one Harris supporter. Here's what the rest had to say:

"Illegals — they gotta go."

"I'm extremely… pic.twitter.com/ThxtVJDEJH

— Charlie Kirk (@charliekirk11) September 17, 2024

Then there are union voters, who broke toward Biden significants, and they seem to hate Harris with a passion. That's a huge constituency, especially in union states like Pennsylvania where the election will be decided. 

YESTERDAY: Nearly 60% of @Teamsters are voting for President Trump.

TODAY: 70% of Steamfitters Local 638 are voting for President Trump.

ALSO TODAY: It’s estimated 65-70% of UAW members are voting for President Trump.

Hardworking Americans know President Trump HAS THEIR BACK. pic.twitter.com/fFOrRs9aXd

— Byron Donalds (@ByronDonalds) September 19, 2024

Harris simply isn't racking up the numbers she needs in crucial constituencies, and the only one where she will outperform is likely younger women. 

White Dudes for Harris aren't going to carry her across the finish line. 

So, while it is wrong to ignore the polls, it isn't wrong to distrust them. Not because they are partisan leaning--although I believe many are--but because their modeling voodoo includes something akin to eating the cats and dogs. (Just kidding.) 

In other words, their model of the electorate is almost certainly wrong. 

Of course, much of the evidence for my own judgment and Professor X's is based on anecdotal evidence or anecdata, and hence also based on questionable assumptions, I wouldn't rely on it. 

If you are interested in the horserace, follow lots of different modelers and pick your poison.