Профиль

Updated dashboards with more non-prob panels.

Time for Congress to get creative about enforcing its powers.

A box of scraps.

I don’t really disagree, but I’ll say as someone who talked to a lot of them while writing the book: the generational change in ability is staggering.

New Heritage staffers are “More MAGA than Trump.”

A smaller but still important gap has opened up over the last year between polls with live interviewers and polls conducted with address-based online probability samples.

Probability sampling = everyone in the sampling frame (ideally target population) is eligible to be included in the survey. Classic probability sampling is random digit dialing, where you just call 1000 random phone numbers and follow up if they don't pick up. Non-prob sampling = opt in in some way

Added Echelon, TIPP, HarrisX non-probability panels. Look at the consistency of the y-intercept on approval.

p<0.05 though

Pretty much. Texas was R+6.5 in 2020 when the country was roughly D+4. Alaska, Ohio, Iowa a hair closer but same ballpark. All easily within the range of normal candidate effects.

D+8

Just YouGov and RMG here.

Sure

@gelliottmorris.com

Of course, a huge typo in the first post.

I've posted these dashboards here. I may or may not update them going forward. www.ejfagan.com/polling

Some thoughts. First, if I were running a real polling aggregator, I would exclude non-probability samples entirely. This choice erases most of the weighting variation. The top line doesn't change a ton in my dashboard, but I only included two non-probability samples. Others include many more.

Okay, now weighting methodology. Smaller but consistent differences here as well. Unlike sampling, I don't have any real opinion on the best choices. Trump/Rs do 2-3 points better if you weight for Party ID or recalled 2024 vote. Some of this variation is correlated with sampling methodology.

I'm only using relatively high quality non-probability online panels here. I'm sure that a lot of generic ballot polls use cheaper, lower quality samples. Note that "lower quality" samples actually performed better in 2020 and 2024, but I would chalk that up to random data rather than good data.

But the real story are methodological differences. Let's start sampling. There's a pretty big difference in both series by methodology. Non-probability online panels are strongest for Trump/R by a large margin. On the generic ballot, phone RDD-ish polls are ~D+8 versus ~D+4 for non-prob.