It seems Joe Biden didn’t like what recent polls were telling him.  So he and his staff started claiming they’re fake.

Unfortunately for Biden sycophants, the most recent Rasmussen Poll had Trump up by a gob-smacking +12 in a five-way race.  Since it is tough to put a positive spin on something like that, and since there is no such thing as a silent Biden voter, this poll had to be wrong, too.

One of the most interesting things about this particular poll is that it is included in the Real Clear Politics polling average but not in the FiveThirtyEight polling average.  Why?  Because Rasmussen was banished from the FiveThirtyEight average in March of this year for supposedly being inaccurate, or perhaps it was too right-wing or too Trumpy to suit FiveThirtyEight’s master, ABC News.  Speaking of accuracy, FiveThirtyEight is the same polling organization that had Joe Biden up by over 8 points in Wisconsin in 2020, which Joe won by less than a point.

For reference purposes, the Media Bias/Fact Check website claims that Rasmussen is a right-of-center organization.  No big surprise there.  Media Bias claims that Rasmussen is guilty of “mostly factual” reporting, which is not high praise.  But it also said Rasmussen has “high credibility.”  Go figure.

If you think that is bad, according to Media Bias, Fox News is the left hand of the devil.  Fox is considered extreme right-wing, with “mixed reporting,” and should be regarded as a “questionable” source of information.

By comparison, the New York Times is seen as left-of-center, with a “high level” of reporting, and as one of the most reliable sources of information ever.  

After ABC demanded that Rasmussen prove that it was worthy of inclusion in their polling average by divulging its proprietary polling methods, Rasmussen told ABC to pound sand.  In response, FiveThirtyEight told Rasmussen to get lost.

Presently, Trump is up in the FiveThirtyEight average by +1.5 percentage points.  So, yes, including the Rasmussen poll would have pumped that average up a bit.

Why do I agree with Biden?  Because polls are statistical amalgamations that contain all kinds of errors.  Call them blurry snapshots in time if you like, with some blurrier than others.

For example, in the Real Clear Politics poll average, the Quinnipiac Poll has Biden at +3 with a statistical sampling error of 2.6%.  If you add the error, Biden could be close to +6 points ahead.  Biden would still be ahead by less than half a point if you subtract it.

Same thing with the Rasmussen survey.  Here, the sampling error is 3%.  If you add the error, Trump could be +15 points ahead.  Subtract, and he is still up by +9.

However, the overall error may be much larger than the statistical sampling error.  The polling results of the 2020 election, where Joe Biden’s support was overestimated by roughly 4% nationwide, suggest something more than sampling error was involved.

Yes, people sometimes lie when responding to a poll.  But it is usually not deliberate.  These lies are produced by social desirability bias.  This is where people alter their responses to appear in a good light to the pollster.  For example, suppose you are going to vote for Trump.  You might tell the pollster you are undecided to avoid any perceived controversy.

According to an article from Pew Research, it is thought that the inaccuracy was due to a combination of voter nonresponse, coverage, and mismeasurement errors.  In every case, these errors reduce the ability to get a representative sample of voters.

When all these sources of error are accounted for, the combined error could be roughly double the sampling error.  So if the sampling error is around 3%, the total error may be more like 6%.

Let’s look at two Reuters/Ipsos polls from earlier this year.  One is from 01/22/24, where Trump is up by +5 with a 3% error.  The second is from 04/03/24, where Biden is up by +4 with a 4% error.

First, if we apply the sampling error of 4% to the 04/03 Ipsos survey, the actual result could be anywhere from a tie to Biden +8.  If we apply a 6% error, as Pew suggests, the result is likely between Trump +2 and Biden +10.

If we do the same for the Ipsos 01/22 survey, with a margin of error of 3%, we get a result between Trump +8 and Trump +2.  If we double the error to 6%, we get Trump +11 to Biden +1.

This means that the envelope of within-the-margin results could have been anywhere from Trump +11 to Biden +10, a range of 21 points, over a ten-week period.  I’m not sure I buy this Ipsos flip-flop, but this shows how much variability there is in polling.

It is not just the individual polls that can be biased.  The polling averages can create their own bias.  Averaging polls is a way to remove bias in polls that are off by random amounts in either direction.  This falls apart if the institution creating the average uses too many like-minded polls.  If the polls are biased in one direction, averaging will not help.

The FiveThirtyEight polling average could fall into that trap.  According to its website, FiveThirtyEight performs three processes to remove bias.  First, it picks its poll buddies using the highest standards.  This means you can’t be on the list if you deliberately mislead the public, follow unethical practices, or don’t kowtow to FiveThirtyEight’s demands regarding polling methods.

Second, it weights every poll by assigning 1 to 3 stars to every polling organization.  Stars are assigned based on FiveThirtyEight’s assessment of a polling organization’s accuracy, honesty, etc., with 3 stars having the most weight.  Unsurprisingly, the New York Times and ABC News/Washington Post are two of the 3-star organizations.  For comparison, the Trafalgar Group got 0.7 stars; CNN got 2; Fox News got 2.8 stars by association with Beacon Research/Shaw & Co.; the Ipsos poll, mentioned above, got 2.8 stars.

Third, FiveThirtyEight performs math magic by fitting the weighted poll data to a proprietary model.  Then it averages them and makes further adjustments to improve the final product.  

There is just one problem.  At least eight polling or news organizations in the FiveThirtyEight average have a left or left-of-center bias.  I suspect that most of the colleges, including Marist, Siena, Quinnipiac, Monmouth, Suffolk, Harvard, and Marquette Law, have a left-of-center bias, too.

I am sure FiveThirtyEight will say that does not matter.  It claims that its averaging process will do an excellent job of removing bias inherent in the polling process.  Perhaps it does.

There is also something quite real in human psychology called confirmation bias.  This means people tend to filter information to fit their own prejudices.  Considering that most of these organizations have a left-of-center bias, is it possible that they are unconsciously designing polls to fit their worldview?

So, yes, I agree with Biden.  The polls may be wrong.  

Trump may be up by more than everyone thinks.

Image: Gage Skidmore via Flickr, CC BY-SA 2.0.