Thread by @aedwardslevy

Here's a first look at the preliminary research into the 2020 general election polling error.

So far, it's most definitive in ruling out potential causes -- e.g. education weighting, late vote shift -- that *don't* seem to have been major factors.

cnn.com/2021/05/13/pol…

This is no surprise if you've been following discussions, but focus is coalescing around some form of differential non-response - that is, people who were polled are different from people who weren't, in a way current weighting didn't account for.

cnn.com/2021/05/13/pol…

For instance, even if a poll included the correct number of Republicans compared to the actual electorate, Republicans in the poll might have been less supportive of Trump than Republicans who weren't reached.

These are preliminary findings -- the AAPOR task force (which, disclosure, I was a part of) is expected to release a full report soon.

In the meantime, important as always to note both that 1) this was a significant error that deserves a lot of further work and 2) polling in general is still not useless or broken beyond repair. See, e.g. recent surveys on vaccines, which line up w/ official data.

Even at its absolute best, polling is always going to give you estimates, not precision. But if you understand those limitations, it can still be a valuable tool for understanding the people around you at a macro level.