Read on Twitter
I wouldn't call it a flub; it's a ~6 pt error more than 3 weeks out!
But it is an interesting question https://twitter.com/BruneElections/status/1304530197555838976?s=20
But it is an interesting question https://twitter.com/BruneElections/status/1304530197555838976?s=20
As a general rule, our polls fared quite well in 2018 but our results in the heavily Hispanic southwest were not nearly so good. Not all of them. But the record there was plainly worse than in the rest of the country, where the average bias was ~0 and average error < 3
In the post-election analysis, the explanation is quite straightforward: we systematically underestimated turnout, and particularly Hispanic turnout. And by a *lot* in some districts.
Our turnout estimates have two components, basically: whether voters have a track record of voting, and whether they say they'll vote. Neither got the job done, at least in 2018
2018 was really tough on the 'track record of voting' stuff, since midterm turnout was wayyy higher than ever before, but especially in CA/TX, where midterms are usually totally noncompetitive and turnout increased by wild amounts. We baked some increase in, but not nearly enough
This was fairly predictable, tbh, as I wrote before live polling again
But as I wrote in the next graf, the hope was that voters would nudge us out of the issue
Truth is, it didn't really happen--particularly among Latinos. Even weeks before the election, low turnout Latino voters just weren't indicating their intention to vote in the kind of numbers that would move our estimates. They were also likelier to be undecided
It turns out, maybe not surprisingly, that our polls in all of these places would have been very good, let alone merely better, if we got the turnout right. We would have had Heller behind in NV with the 2018 electorate, for ex.
Now, I'm not saying we should get credit for showing a Heller defeat. But for a pollster, there's a big difference between getting a group wrong because you're sampling them poorly or because your electorate is wrong.
If it's sampling, then we'd be persistently biased against Democrats in the Southwest. Turnout, on the other hand, can go either way from race to race. We could overestimate Latino turnout in 2022 if Biden wins and be biased toward Dems in the exact same spots.
For all we know, we'll be biased toward Dems in these same spots in 2020. I'd argue the two AZ polls we've done so far were unusually strong for Biden, v other firms, at those points of the race. What if we have a Dem bias there obscured by high turnout in '18? IDK
