Looking at the 2020 election, what conclusions should we draw about poll analysis sites like 538, etc.?

Nate Silver gave Biden a 90% chance to win and Biden did win. He even indicated that it was difficult to tell if it would be a blowout(what I expected) or a nail-biter(what we got). Obviously, it was a huge blowout from a popular vote POV, but the EC really scared us.

Was there a polling error similar to 2016? Was it worse or less bad? Are polls having a hard time determining who will actually vote?

Is it useless to utilize poll-analysis anymore?

Let’s wait until the count is finished. I expect we’ll find that the polls were dead on in some states (like GA), and off by a small to moderate amount in some others. But we won’t know for sure until we get final numbers.

The error was bigger than 2016 I think. Especially in the rust belt, 2016 Hillary was supposed win by something like 2% and ended up losing by a rounding error. In 2020 Biden was supposed to win by around 8% and ended up winning by 1% in Michigan and less in Wisconsin and Pennsylvania.

It’s possible this particular election just bucked all the trends among likely voters because there was so much mail-in voting, but I think polling is getting less accurate. I wonder how much 538 updates what they look at for good polls after this.

I don’t know if anything could be done. I was very forgiving after 2016. I expected Nate and the others(this isn’t just about 538) to make major adjustments…and most said they did make major adjustments to their polling and poll analysis.

Still wrong. I’m from Michigan and we just elected a Democratic governor in 2018 by quite a large margin. Even I expected Biden to win by 5% or so, maybe more. Nope. Close election, same mess.

Speaking as someone who has worked in market research for much of his career, I think that opinion polling is kind of a broken science right now.

It’s not really the pollsters’ fault, either. It’s becoming harder and harder to get a truly representative sample in any sort of consumer research – a lot (probably most) people won’t answer a phone call (or a text) if it’s not from someone they know, and the prevalence of spam in those media makes people less willing to trust that an unsolicited caller is who they say they are. Online surveys aren’t any better or any more representative, either, and there is an increasing percentage of Americans who simply don’t want to be bothered by a surveyor/researcher/pollster.

I also think that there’s a segment of the population who is now pretty unwilling to trust any mass media/news outlet, so if they get a pollster calling from “ABC News” or “the Washington Post,” they are likely to just hang up (and a segment of them may choose to instead lie to the pollster, just to mess with their numbers). I also suspect that the people who feel this way are conservative-leaning, meaning that their opinions are being underrepresented in poll results.

Finally, I do also think that there is a certain, small percentage of “shy Trump voters” – people who will vote for Trump (or similar candidates), but have felt scorned by others for that view, so they simply won’t talk about it or admit to it, even in a theoretically anonymous poll.

I’m not sure that any of that can really get fixed any time soon, either.

538 was much more cautious this time around, and I felt did okay. They’ll Have a lot to say on this subject this week. The margins were less than predicted but it wouldn’t surprise me if Biden ends up winning the popular vote by 7m. Ohio and a few senate races were the biggest surprise - the fact it came down to the WOMEN was not. (My coinage! Wisconsin, Ohio, Michigan, ‘Ennsylvania and North Carolina).

Yes, Nate says 81 million to about 74 million. It was a huge blowout. In any other nation, this would have been an obvious and clear result pretty quickly.

Your numbers are off, and they’re still counting. Biden is up in Michigan by a few points right now.

I always get cognitive dissonance when people say things like that. I kinda get that it is supposed to mean something like “in the context of closely fought modern elections, this represents an unusually high repudiation of candidate X.

But I can’t really accept that particular context. A huge blowout TO ME would have been Trump dropping even just a bit relative to his last performance and Biden scooping up all the extra. Say Biden ~89-90 million and Trump ~62 million.

This wasn’t a huge blowout. A huge blowout is Reagan vs. Mondale. It’s not even a little blowout. This was a closely fought race with yes, a clear winner. It’s a lot less close than the Trump victory of 2016. But it isn’t a crushing victory, it’s a fairly close one. More’s the pity, because I was really hoping for a crushing victory and I am already more than a little concerned about 2024 :roll_eyes:*.

*The roll-eyes are for me. I really shouldn’t be thinking about 2024, until at least, oh, February of '21 :wink:.

It looks like the results (which aren’t quite in) will fall in the middle of Silver’s wide band of most likely results, for President, Senate and House. Silver’s estimates were more cautious than a naive review of the polls would have come up with - and I think the results justify the caution.

I agree. The presidential results were well within the margin of error. The only swing state that didn’t quite match was FL (what else is new?).

Some of this analysis is preliminary and subject to change if any states flip from what they look like now due to recounts or court cases. This assumes Biden will win the EC 306 - 232

538 correctly predicted 53 of the 56 electoral vote awarding entities (~95%). He only missed Florida, Maine CD-02, and North Carolina. He had Biden’s probability of winning each of those at less than 51%.

All of that is pretty good. What’s making him look bad is that all his misses were in the same direction so his modal EC outcome 351 - 187 looks pretty far off. But the actual outcome was right in the fat part of his projected distribution of all outcomes.

All that said, PredictIt ate 538’s lunch.

PredictIt was right in 54 of 56 EC awarding races. It only missed Georgia and North Carolina, missing one in each direction. The modal EC outcome per PredictIt was 305 - 233.

Here’s a slope chart of PredictIt probabilities (debiased for favorite/longshot bias) compared to 538 on election for races where either gave any party a 10% or greater chance to win. There were 18 such races and surprises outside these 18.

Google Photos

538 was more Biden leaning than PredictIt pretty much across the board.

Here’s that same info in table form with a column for results.

Race EV PredictIt 538 Result
SC 9 0.004 0.100 0
AK 3 0.015 0.149 0
MT 3 0.027 0.180 0
IA 6 0.140 0.399 0
OH 18 0.162 0.453 0
TX 38 0.238 0.382 0
GA 16 0.383 0.582 1
FL 29 0.417 0.691 0
ME-02 1 0.520 0.573 0
NC 15 0.561 0.639 0
AZ 11 0.645 0.680 1
PA 20 0.749 0.843 1
NE-02 1 0.837 0.717 1
MI 16 0.849 0.947 1
NV 6 0.876 0.878 1
WI 10 0.887 0.944 1
MN 10 0.897 0.958 1
NH 5 0.910 0.891 1

If you calculate Mean Squared Error across these 18 races, you get:

PredictIt → 0.0849
538 → 0.1225

Lower is better.

If you plot MSE error of both over time across al races you see that while both improved as the election drew closer, 538 was always worse. In fact, 538 in November was worse than PredictIt at any point going back to mid-April.

Google Photos