Thread by @baseballot

🧵: We looked into the TIPP/American Greatness poll of PA that has been making the rounds on here. If you’re not familiar, you can find the poll here: amgreatness.com/2024/10/10/tru…

The poll got our attention because, while its registered-voter version included 124 RVs from Philadelphia, 116 of whom said they were “very” or “somewhat” likely to vote, the LV version had only 12 LVs from there.

We thought this might be an error, so we reached out to TIPP & American Greatness. The president of TIPP responded and said it wasn’t an error: This poll’s LV model assigns a vote probability to each respondent based on their demographics & vote history, and…

… a disproportionate # of their Philly respondents had factors that made them unlikely to vote: they were young, not a college graduate, and/or nonwhite. So a ton of them happened to fall beneath the threshold to count as a likely voter.

(He also said that respondents’ answer to the “how likely are you to vote” question was a factor in this poll’s LV model as well, but only a small one. He also confirmed that TIPP alone, and not American Greatness, determined the LV model.)

Now, this is a different LV model than some past TIPP polls have used. For example, their September polls of AZ and NC counted everyone as a likely voter if they reported that they were “very” or “somewhat” likely to vote. (But, notably, more recent GA and NV polls did not.)

When I asked about this, he said that TIPP actually has 5 different LV models. While he would not tell me how they decide which LV model to use on which poll, he did say that they have decided to use the PA model on all polls going forward.

What’s the bottom line here? I’m personally not putting a lot of weight on the PA poll. It seems like it ran into a bad sample that the LV screen made worse. IMO, ideally you’d weight the LV sample by region, but he told me they only weight the RV sample.

But at 538, we don’t police pollsters’ LV screens. (As a reminder, appearing on our polls page is not a guarantor of quality, just of existence.) The poll meets all standards in our policy (), so we include it.abcnews.go.com/538/538s-polls…

If you’re worried about it affecting our models, don’t be. We weight polls by quality and account for “house effects,” a.k.a. bias. To give you a sense, this poll moved our PA average all of 0.1 percentage points, from Harris+0.7 to Harris+0.6.