A failure in representative polls like this should be avoided with statistical methods.
The margin in Pennsylvania will continue to shrink, as the only place with lots of votes left to count is Philadelphia. Michigan might still flip blue, because the only place with votes to count is Detroit. Arizona is still a total coin toss, with 51k vote difference and >1200k votes left to count. Wisconsin is going to be close too, although it will likely stay red.
None of that matters when there are less ballots left to count than the margin in PA, but still, the message from the polls before election was "this will be a nailbiter", and it kind of was.
What they said was that they could not predict the outcome, and were giving basically 50/50 odds of either candidate winning, which is essentially just another way of saying "I have no idea".
Just because their odds were 50/50 though, does not mean the outcome would be close. The pollsters were all warning that the swing states would likely be strongly correlated, so if a candidate performed strongly in one swing state, they'd probably perform strongly in all of them.
My explanation for this is that most polls were fabricated, showing enthusiasm for (D) which wasn't there. Basically, a form of propaganda. The most striking example here is Selzer, with that Harris+3 Iowa poll the day before the election.
In reality, a lot more people have traditional values when it comes to race, LGBT whatever, sexism, spiritual values, opinions on Russia, Israel etc. However in public they may be scared to voice their true opinions.
Use Polymarket instead, where money is on the line.
If I predict a coin toss to be 50/50 that doesn’t mean I expect it to land on its side.
The people in a swing state choosing to spend time responding to polls are insufficiently representative. They're drowning in advertisements, calls, texts, unexpected people at their door and randos on the street. Why would they give time to a pollster?
For example, from Nate Silver's blog:
> The Silver Bulletin polling averages are a little fancy. They adjust for whether polls are conducted among registered or likely voters and house effects. They weight more reliable polls more heavily. And they use national polls to make inferences about state polls and vice versa. It requires a few extra CPU cycles — but the reward is a more stable average that doesn’t get psyched out by outliers.
All this weighting and massaging and inferencing results in results that are basically wrong.
Come Election Night he basically threw the whole thing in the trash too!
It is like when your doctor is asking you if you eat fast food — some people will downplay it because they know it is wrong, but do it anyways in a "weak" moment when nobody is looking.
So suddenly in my village where I know everybody 56% voted for the right wing candidate, yet everybody¹ claimed not to do that when asked before or after.
¹: except one or two open Nazis
Messaging is build on focus groups, and tweaked to get the best results by both sides. That group is the same group that does polls.
Its a Goodhart's law in action: Any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes.
Polling companies are in the business of media deals and government contracts. They will develop methodology and reporting to that end and the money is in "a close and contested race", even if it won't be.
(and this was while Biden was still in the race)
Pollsters such as Nate Silver were giving gut-takes of Red over Blue, e.g.:
"Nate Silver: Here’s What My Gut Says About the Election, but Don’t Trust Anyone’s Gut, Even Mine" (Oct. 23, 2024)
<https://www.nytimes.com/2024/10/23/opinion/election-polls-re...>
I've done a somewhat half-assed take tonight of comparing actual returns to latest pre-election polling by state
Why that is, isn't clear. Political pollsters have been struggling for years with accuracy issues, particularly as landline usage falls (it's <20% in most states now), and unknown-caller blocking is more widely used (both on landlines and mobile devices).
Polling does have periodic calibration events (we call those "elections"), but whatever biases the polls seem to experience in the US, it's apparently systemically exceeding adjustment factors.
Polls / votes and deltas:
QC State EV BP RP BV RV Bd Rd
4: AL 9 36 64 32 65 -4 1
4: AK 3 45 55 0 0
4: AZ 11 49 51 49 50 0 -1
4: AR 6 36 64 34 64 -2 0
4: CA 54 63 37 60 37 -3 0
4: CO 10 56 44 55 43 -1 -1
4: CT 7 59 41 54 44 -5 3
4: DC 3 92 7 90 7 -2 0
4: DE 3 58 42 56 42 -2 0
4: FL 30 47 53 43 56 -4 3
4: GA 16 49 51 48 51 -1 0
4: HI 4 64 36 0 0
4: ID 4 33 67 33 64 0 -3
4: IL 19 57 43 52 47 -5 4
4: IN 11 41 57 39 59 -2 2
4: IA 6 46 54 42 56 -4 2
4: KS 6 42 51 41 57 -1 6
4: KY 8 36 64 34 64 -2 0
4: LA 8 40 60 38 60 -2 0
4: ME 2 54 46 0 0
4: ME-1 1 61 39 0 0
4: ME-2 1 47 53 0 0
4: MD 10 64 36 60 37 -4 1
4: MA 11 64 36 62 35 -2 -1
4: MI 15 50 49 0 0
4: MN 10 53 47 0 0
4: MS 6 40 60 37 62 -3 2
4: MO 10 43 57 42 56 -1 -1
4: MT 4 41 59 33 64 -8 5
4: NE 4 41 59 42 56 1 -3
4: NE-2 1 54 46 0 0
4: NM 5 54 46 51 47 -3 1
4: NV 6 50 50 0 0
4: NH 4 53 47 52 47 -1 0
4: NJ 14 57 43 51 46 -6 3
4: NY 28 59 41 55 44 -4 3
4: NC 16 49 51 48 51 -1 0
4: ND 3 33 67 31 67 -2 0
4: OH 17 46 54 44 55 -2 1
4: OK 7 33 67 32 66 -1 -1
4: OR 8 56 44 55 43 -1 -1
4: PA 19 50 50 0 0
4: RI 4 58 42 55 42 -3 0
4: SC 9 44 56 40 58 -4 2
4: SD 3 36 64 29 69 -7 5
4: TN 11 38 62 34 64 -4 2
4: TX 40 46 54 42 57 -4 3
4: UT 6 39 61 43 54 4 -7
4: VT 3 67 34 64 32 -3 -2
4: VA 13 53 47 51 47 -2 0
4: WA 12 59 41 58 39 -1 -2
4: WV 4 30 70 28 70 -2 0
4: WI 10 50 49 0 0
4: WY 3 73 27 70 28 -3 1
Blue votes: 43
Red votes: 43
Blue delta: -2.49
Red delta: 0.63
- QC: A parsing QC value (number of raw fields)
- State: 2-char state code, dash-number indicates individual EVs for NE and ME.
- EV: Electoral votes
- BP: Blue polling
- RP: Red polling
- BV: Blue vote return
- RV: Red vote return
- Bd: Blue delta (vote - poll)
- Rd: Red delta (vote - poll)
The last two results are the cumulative average deltas. Blue consistently performed ~2.5 points below polls, red performed ~0.6 points above polls.
Data are rounded to nearest whole percent (I'd like to re-enter data to 0.1% precision and re-run, though overall effect should be similar). Deltas are computed only where voting returns are >0.
Data are hand-entered from 538 and ABC returns pages.
Blue consistently polled slightly higher than performance. Polls don't seem to include third parties (mostly Green, some state returns include RFK or others).
There are all but certainly coding/data entry errors here, though for illustration the point should hold.
That pool was apparently more the former than the later.
https://www.realclearpolling.com/polls/president/general/202...
It was no surprise he won, IMO.
Also note that a "90% / 10% change to win" is not necessarily "wrong" if the 10% candidate wins. Anyone who has played an RPG will tell you that 90% chance to hit is far from certain. Maybe if there had been 100 elections, Clinton would have won 90 of them.
Nate Silver nailed this in the 2016 election. He said Trump's victory there was consistent with historically normal polling errors.
What may have been less widely appreciated is these errors are not related to causes like limited sample size that are straightforwardly amenable to statistical analysis. They come from the deeper problems with polling and the way those problems shift under our feet a little bit with each election.
Trump won many of those states by 2-3%.
Otherwise the real result would be distributed around the mean within the margin of error.
There is some bias and the polls did not correctly factor that into their statistical model.