POLLS EXPLAINER: it's natural to want to compare 2020 election polls to 2016's, especially in swing states, but here's why that might not be such a good idea. My story in the @FT today reviews the key differences: (THREAD)
ft.com/content/b32976…
ft.com/content/b32976…
1/ there's a lot of talk abt Trump/Biden betting odds being at 50:50, but here's a chart via @martinstabe showing in 2016, betting mkts were confident Clinton would win. in surveys more ppl including Trump voters also said they thought Clinton would win ft.com/content/3c9487… 
2/ this yr it's the opposite perception; more ppl/mkts think Trump will win. despite Biden's robust national polling lead, compared to Clinton's. now let's address the "the polls were wrong" critique. it's an understandable reaction to 2016, but *national* polls did pretty well!
3/ quote from an @AAPOR research study: 2016 national polls were "among the most accurate in estimating the popular vote since 1936" ... OK, you might say, but who cares about "estimating the popular vote"?
aapor.org/Education-Reso…
aapor.org/Education-Reso…
4/ see: 2000 & 2016 when Dems won pop vote & lost White House. In 2016, it's true that *state* polls were off; in Wisconsin they were off by an avg of 6.5 points, in PA & NC it was 3-4 points, skewed toward Clinton
3/ TBF, a few pts is not "a LOT", but in a close election it really matters, esp as they tended to underestimate Trump. but worth mentioning that if your position is "the (state) polls were wrong in 2016", you should think twice abt using those same polls in a comparison to 2020
4/ why were '16 state polls off in one direction? one issue is many didn't weight by education. highly-educated ppl are more likely to respond to surveys. this @Muhlenberg @mcall poll of PA voters would have cut its '16 error in half if it weighted by education, which it now does
5/ this isn't an indictment of pollsters, btw. one way to think about survey weights is that you weight on a variable if (a) it is correlated w/response rate AND (b) it is correlated w/the outcome you're studying (voting patterns). in the past only (a) was by&large the case
6/ Three pollsters who now weight by education — Chris Borick of @Muhlenberg, @mbpolisci of @UNFPORL, & @MikePNoble of @OHPredictive — were kind enough to re-calculate their latest 2020 numbers w/o out the education weighting. Biden's lead looks larger as a result.
7/ They didn't weight by education in 2016, so if you wanted to compare their 2016 state polls to 2020's you'd need to mentally add a few pts for Biden or subtract a few pts for Clinton. Again, a few pts isn't "a lot" but neither were the razor-thin margins in key states in '16
8/ So does that mean you should just not compare '16 state polls to '20 at all? Not necessarily. There are some useful analyses like adding up %Biden+%Trump and seeing how many voters are left "up for grabs." Those %s are much smaller this year
9/ That's not to say I don't think the race is tightening! Or that I can predict November! Rather, I hope this story is a useful guide to explain the current state of polls, and why/how things have changed since 2016 ft.com/content/b32976…
10/ Sidenote this was my first solo-bylined story for the FT, pls read it so it won't be my last 😅
HT @JoshTANoble @MattGarrahan for the idea, @bydavidcrow for making my words make sense + @jburnmurdoch @martinstabe @joannaskao @LaurenFedor @federicacocco for #dataviz advice!
HT @JoshTANoble @MattGarrahan for the idea, @bydavidcrow for making my words make sense + @jburnmurdoch @martinstabe @joannaskao @LaurenFedor @federicacocco for #dataviz advice!
