I'm not sure the Locals crowd needs this to be said, but I'm going to repost here what I wrote on X about this topic, with a bit more detail.

It started when Sean Trende on RCP attempted to warn people about modeling the independent vote solely using polling data. His points, particularly his main point, is very true. Here's what he said:

"If you're trying to model "Independent" voting, polls are of limited use. Remember that depending on question wording, context, and pure randomness, "Independent" can mean (among other things): Registration, how people think of themselves, or voting tendencies."

Original Post URL: https://x.com/SeanTrende/status/1852718833548620274

More to his point, how someone refers to themselves by party when speaking to a pollster or an exit pollster can be entirely different than their actual voter registration record.

It's how Trump, for instance, can win Hyde County, North Carolina, when it was majority Democratic in registration. The same is true for most of the 31 counties that voted for Obama, then Trump in 2016 and again in 2020, as well as cause of sharp swings in Appalachian Western Pennsylvania and Northeastern Pennsylvania counties.

If you polled a registered Democrat in Hyde County, NC, who supports Trump, they're very likely to tell you they're an independent, rather than identify with their actual registration. Voter registration is more static and lags, changing over time to follow more consistent and fluid trends in party identification.

It's very tricky to use polling data ONLY to model "indies" if the pollster didn't data append to the voter file, as we almost always do, so they know party registration.

Here, let me show you how this can be observed in a real situation.

The following is from a poll that we conducted in Utah and it was appended, otherwise known as matched or deduped, to the voter file.

The first image is the self-identification question, the one you're all familiar with from any poll you see on Real Clear Politics (RCP).

The second image is the result of the data append, or again, matching to the voter file.

As you can see, when polled 43.2% of the sample self-described as Republican, 31.3% as Independent, 18.0% Democratic and 7.4% as Something Else.

But the data append showed that the registration record by party for the very same sample was actually 50.9% Republican, 17.1% Democrat, 14.9% Independent American, and only 12.2% Unaffiliated, etc.

When people talk about polls, they always assume Independent means Unaffiliated, and visa versa. That is not really the case. Sometimes we get lucky, but often we do not. It's yet another reason why I stress "sampling errors for subgroups are higher" as much as I do.

But the bottom line, a lot of these micro "models" floating around attempting to project the margins using polling data by independent vote ALONE are not really better than a good guess.

In reality, it's far more complicated than that, unfortunately for me. Other demographics tell us about who these people are and how they're likely to vote regardless of having no two-party registration, such as age, race, sex, etc.