George ConwayOct 26, 2024
🧵: The fact that “All polling is based on modeling from past elections,” which Stuart accurately notes, is something most people don’t understand.

It’s a necessity of public-opinion research. Because if you call up 100,000 people randomly, you’re going to get a good sample of people who answer their phones—not a good sample of the adult population generally, or the pool of registered voters.

RE: https://www.threads.net/@gtconway3/post/DBl002ux4fo
George Conway (@gtconway3) on Threads

Threads
Show threadGeorge ConwayOct 26, 2024
And you certainly don’t necessarily end up with a sample that adequately represents the people who will actually cast votes—because the identity of those people can’t be known until after they vote. (Also worth noting is that if you randomly dial 100,000 numbers these days, you might only get a few hundred responses—and that doesn’t help, either.)
Show threadGeorge Conway
You can ask people “are you gonna vote”—and you’ll get a lot of people who say they will vote but may not. It’s like asking, are you going to go to the gym today. “Sure I will.” Oops—it’s about to close; maybe next time. You can ask if they voted in the past, but humans are not always consistent.
Oct 26, 2024 at 2:46pmWeb
Show threadGeorge ConwayOct 26, 2024
So pollsters have to take whatever sample they have and weight the responses (by age, gender, party affiliation, race, whatever) to approximate their educated *guess* about the demographics of the people who ultimately cast votes. But those demographics are not knowable until after everyone votes.

Derek Jeter used to say about baseball, “this is why we play the games.” Elections are much the same.
Show threadGeorge ConwayOct 26, 2024
By the way, all of the foregoing explains why polls are really good for two things—message testing and identifying trends—and less so for predicting outcomes in relatively tight races.