Experts confounded by 2020 polling error

[pilight]

https://theconversation.com/survey-experts-have-yet-to-figure-out-what-caused-the-most-significant-polling-error-in-40-years-in-trump-biden-race-160967

Quote

 

Their findings did make clear, however, that the 2020 miscall was the most significant in 40 years.

Polls in the presidential race in 2020 collectively overstated Biden’s lead by 3.9 percentage points, the task force chair, Joshua Clinton of Vanderbilt University, said in a presentation at the conference.

This marked the fourth presidential election in the past five in which the national polls, at least to some extent, overstated support for Democratic candidates.

 

 

[vcczar]

7 minutes ago, pilight said:

https://theconversation.com/survey-experts-have-yet-to-figure-out-what-caused-the-most-significant-polling-error-in-40-years-in-trump-biden-race-160967

 

I think part of the issue this year was Covid and mail-in voting, etc. It made everything harder to predict, even though I predicted the EC almost exactly (missed ME-2 only). 

I think Trump support is hard to estimate. Probably a lot of them don't respond. I also think the decrease of LAN lines is making polling more difficult, because people don't tend to pick up unknown calls from a cell phone. Polling is polling support from those that will take polls.

[Pringles]

I think we have a "Shy Tory" effect here in the states. 

Or Bradley Effect, although I think that's a bit different. 

I think from now on every time I take a look at polling, unless the Democrat has over a 3% lead at least I'm considering it a to close to call kinda deal haha. 

[vcczar]

8 minutes ago, Pringles said:

I think we have a "Shy Tory" effect here in the states. 

Or Bradley Effect, although I think that's a bit different. 

I think from now on every time I take a look at polling, unless the Democrat has over a 3% lead at least I'm considering it a to close to call kinda deal haha. 

I usually assume a 2% exaggeration for Dems in states controlled by a Republican legislature because of voter suppression. This helps me predict the 2020 election.

[Rezi]

I think our polling problem is less a methodology problem, and more of a Trump problem. Polls since 2016 have been fairly accurate when Trump isn't on the ballot (see 2018 midterms, Georgia Senate Runoff), but everytime Trump is on the ballot they way overestimate. I think that Trump just turns out a different kind of voter that wouldn't vote without him or a figurehead like him there.

[Timur]

Trump is one of those candidates that has many silent supporters.

[Entrecampos]

22 hours ago, vcczar said:

I think part of the issue this year was Covid and mail-in voting, etc. It made everything harder to predict, even though I predicted the EC almost exactly (missed ME-2 only). 

I think Trump support is hard to estimate. Probably a lot of them don't respond. I also think the decrease of LAN lines is making polling more difficult, because people don't tend to pick up unknown calls from a cell phone. Polling is polling support from those that will take polls.

I made a politics/economic model mixed with an elo pollster rater that was right in all states by 4% margin. I predicted 3.8% advantage for Biden Only instead of 4.5% and called 3 wrong states (Wisconsin, Georgia and Arizona), but all could be right if my national vote advantage for Biden was right. And would be if I knew the real gdp 2020 growth annd unemployment rate like we know now.

I think Bolsonaro polls have the same problem. Today was out one where he would lose 66-34,but I remember when he had 30% in first round 3 days before 2018 election in some polls and had 46%.

Edited May 21, 2021 by Entrecampos

[vcczar]

5 hours ago, Entrecampos said:

I made a politics/economic model mixed with an elo pollster rater that was right in all states by 4% margin. I predicted 3.8% advantage for Biden Only instead of 4.5% and called 3 wrong states (Wisconsin, Georgia and Arizona), but all could be right if my national vote advantage for Biden was right. And would be if I knew the real gdp 2020 growth annd unemployment rate like we know now.

I think Bolsonaro polls have the same problem. Today was out one where he would lose 66-34,but I remember when he had 30% in first round 3 days before 2018 election in some polls and had 46%.

I wasn’t surprised when my algorithm was showing WI and AZ for Biden. It seemed obvious for weeks. However, GA surprised me. Glad I let my algorithm do the speaking for me. My only miss was ME-2. There wasn’t that much polling there. 

[Entrecampos]

20 minutes ago, vcczar said:

I wasn’t surprised when my algorithm was showing WI and AZ for Biden. It seemed obvious for weeks. However, GA surprised me. Glad I let my algorithm do the speaking for me. My only miss was ME-2. There wasn’t that much polling there. 

Wisconsin wasnt as obvious as polls were showing, region is turning more GOP since 2012 elections. Historical polling in Wisconsin usually is horrible. 2018 midterm elections showed GOP keeping all CDs (I didnt included that info in my predictions altought). I had strong arguments for a more disputed election there. And mainstream pollsters were the worse of all in general. When I seen +17 Biden Wisconsin in one poll I started joking. Because my model was only 30% based in polls, I could understand that Trump would do better in Wisconsin than predicted.

Edited May 22, 2021 by Entrecampos

[Dobs]

14 hours ago, vcczar said:

I wasn’t surprised when my algorithm was showing WI and AZ for Biden. It seemed obvious for weeks. However, GA surprised me. Glad I let my algorithm do the speaking for me. My only miss was ME-2. There wasn’t that much polling there. 

I got ME-2 but I missed GA. Got everything else correct.

[vcczar]

4 hours ago, Dobs said:

I got ME-2 but I missed GA. Got everything else correct.

That's kind of what I was expecting, but my algorithm told me differently.  I felt 80% confident Biden would win WI and AZ. I was 90% confident of MI and PA going for Biden. I was 40% percent on GA, but my algorithm said it would go for him. ME-2 and NE-2 were the two places I didn't get as much data, so the algorithm went by the few polls that existed and data that was far more incomplete than statewide info I used.