Well... I get that you want to quibble over words, and "lean left" is not strictly scientific.
However, in 2020, 17 of 20 pollsters tracked by Real Clear Politics (RCP) overestimated Biden's vote percentage -- the worst of them, including CNN and Quinnipiac, by over 7%.
People naturally see this as "leaning left" and can even accuse such pollsters of conducting "push polls" hoping to influence the election with their skewed polls.
But another way of looking at it is that Biden hugely underperformed what some pollsters expected, meaning the worst of them were calling for Biden to receive nearly 6 million more votes than he actually did.
Chose your own perspective on that.
In 2020, only three of 20 pollsters came close to the actual results -- Insider Advantage, Susquehanna and Trafalgar -- and they all "leaned right", meaning they predicted a much closer popular vote.
The fourth closest was Rasmussen, and they "leaned left" predicting a slightly larger Biden win than the actual turned out to be.
In the 2022 midterms, everything changed, with most of the top 2020 pollsters -- i.e., Insider Advantage, Rasmussen, Trafalgar -- sliding to the bottom and some of the worst in 2020 -- CNN, FOX, NYTimes -- rising to the top in 2022.
Also, overall, the margins of error were about 1/3 less in 2022 than in 2020, so they all adjusted and came closer to actual.
And returning to the pattern of 2018, in 2022, half the errors were "right leaning", and some were exactly neutral -- CNN, NYTimes & CBS.
So, what will happen in 2024?
I think it's worth noticing that the error pattern of "left leaning" vs. "right leaning" in 2020 was almost identical to the previous presidential election, in 2016, suggesting something happens when Trump is on the ballot which doesn't happen when he's not.
Have all the pollsters made the correct adjustments in 2024, to give us more reliable predictions?
Maybe, but there's no way to know for sure until after the election in November.
Electoral college math suggests that Harris needs more than +3% in the popular vote to win the electoral college and as of today, only five pollsters show Harris ahead by more than +3%:
Take your pick.
Owen: "Which is why these polls are pretty tightly clustered.
None are coming in with 15 point leads in both directions.
They are almost all within MOE."
The spread from Harris +6% to Trump +3% is nine points.
Thank you for the articulate well thought out commentary.
I hope and pray that you are correct in your meta analysis.
My comments were geared toward how easy it is to manipulate insividual survey results by biased sample selection procedures.
Very few people understand the weighted averaging and assumptions utilized to obtain the desired results.
Having been trained in “push polling” as a marketing tool, I understand how easy it is to use a survey to achieve desired results rather than measuring actual population attributes.
That’s a good presentation of history.
But it’s not the way to examine these things.
There has to be a known mechanism and that mechanism is turnout model. The only other way to bias a result would be to exclude neighborhoods from phone calls, and this is just not reasonable when sample size turns out to only be 800ish +/- 300. That is asking pollsters to know neighborhood makeup for not just area code but also phone number prefix. And that fails when some use online surveys.
So the only mechanism of bias has to be turnout model. Declaring this pollster is left leaning vs another that is right leaning does not explain why. How did they achieve the lean. It’s not by phone prefix. It has to be turnout model.
2016 was chockful of analysts (whose paychecks come from running a polling company) explaining loudly how data did not really fail. Data is still great. This was wrong because Trump voters refused to admit it. Well, no, because there is no reason to think Trump voters refuse to admit it in only certain states, and there were indeed states where the polls were accurate.
No. What happened was failed turnout models. The most recent Census was 2010. The party affiliation data was similarly very old. But whereas a pure academic exercise would have announced that there was no possible way to make an estimate, you can’t say that when getting paid.
This year, the Census is only 4 yrs old. The virus corrupted all information in 2020 so largely ignore that. There will be VBM, but the requests for ballots are way down this year vs 2020. And so, do not look at Morning Consult Harris +X. Look at whatever information (often not extensive) you see from Morning Consult on White/Black/Hispanic ratio.
Look, people see this data that says something like Women/Men 53/47 and declare women were oversampled. This isn’t true. Women 53/47 is the correct ratio given they live longer and there are more of them. Racial sampling is how the Harris surge was manufactured. The pollsters decided blacks would turnout more. They generally already turnout 65%+, about the same as whites, but by lifting them 2% to 67% and knowing they vote 80/20 Dem, you can move the needle.
That’s the conclusion. You can’t declare left or right leaning. You have to examine the turnout models. If they justify the model with source of partisan mix and source of man/woman, then it is not they who are biased. It is the source.