The primary use for polls is not pre-election handicapping, it's to understand what the average American wants from their government. And while election polls aren't in a good position, issue polls of all adults seem fine (for now).
I agree that people need to "kick the addiction" of thinking polls are infallible predictors of election outcomes. They're not, that's why we try to model what could happen if they misfire.
Separately, I have noticed this really annoying trend where a lot of the people saying we should "quit" polling are finance or quant folks who want 100% accurate predictions from them, but that's not what polls are and never will be.
I totally agree! There are big problems with the polls and they need to adapt to a dynamic and polarizing environment of social trust in America. But that's different than saying they're "useless" or that they "failed" us or that we should "quit" them.
Some more math. Fulton County alone has 25k absentee ballots left to count. So far, Biden has won 79% of absentee votes there. IF they break similarly, Fulton alone would close Trump's margin to under 3k votes
Similar numbers in Chatham County (Savannah), where the AJC says there are 16k absentee votes to count. I have no way of verifying that, but if they split 74% Biden like the already-counted absentees votes that's a 4k vote net gain for Biden which would push his lead over Trump's
I have heard these "people are not numbers" and "we are not statistics" lines a lot today, and it's unclear to me if any "numbers nerds" are actually arguing this? We pair the majority of our data-driven reporting with interviews with experts or actual voters. It's def not 100-0.
It seems to me that the people being safest with the polling data and who most often note the nuances of the processes that produce the numbers that campaigns and news orgs rely on are the data-driven journalists who actually look at the numbers every day!
It's also pretty funny to see the coalition for innumeracy out in full force today as models are on track to "call" ~50 of 51 states in the pres elec correctly, albeit with some above-avg polling error. Today says more about how J School Types consume polls than about how we do.
#NEW: Polls are exhibiting concerning patterns of geographic bias and partisan non-response that have maybe even gotten a bit _worse_ since 2016. But it is not doomsday for them—or the models. In fact, quantifying uncertainty is more important than ever.
Our models said that Biden had so large a lead, and across many key states, that his position would very likely persist even under even a large systematic polling error. With Biden heading for 290 or maybe even 306 electoral votes, that is... probably exactly what happened.
We have learned a few early lessons about our model.
First, we underrated the chance for error with polls, particularly re: turnout models. We'll use fatter tails next time.
Second, we might think about using a prior of geographic polling error as per 2016-2020.
FOLKS! The initial results of #Election2020 are trickling in from IN and KY. Follow this thread for my analysis and live prediction of results throughout election night.
And remember you can update our election model with your own estimates here ⬇️
If these are anywhere near accurate, Trump is toast.
So, these first results are mostly from absentee votes, so should skew a bit more Democratic than expected. So I am withholding judgement until we get to 100% reporting.
I still can't engage with @NateSilver538 on this site but this tweet misses a v important nuance of the post, which is that a model with fatter tails actually makes our aggregate predictions *worse* for 2008-2016, even though it makes the state-level predictions more inclusive.
This might have something to do with the way we parameterized our a correlated error and how relying on the fundamentals helped us shrink toward 50-50 in 2016 -- Nate needed fat tails then to control for poll error but our model hedged by picking up pro-D bias in swing states.
Also, adding fat tails on the order of what Nate uses only pushes our model down to 93% — so it's not like a huge difference. The bigger reason there's a gap bt our models is that 538 includes some very R-biased partisan data that's pushing their avgs toward 50-50 in key states