This is mostly an example of a kind of survey bias I’m having trouble finding the name for plus a counterintuitive effect of averaging.
For the bias, when people are asked to estimate a percentage, they tend to estimate by large fractions of the whole, like by quarters, fifths, or tenths. This means you’ll see survey estimates closer to 50% than the real value, with the effect more pronounced for real values closer to 100% or 0%.
For the averaging phenomenon, when looking at the averaged responses across all questions of a survey, you can quite easily get a collection that wouldn’t make sense as a set of responses for “the average” (that is, the typical) person. You can have 3 different responders who each think California, Texas, or Florida has more people than they actually do, and then when you average those responses it looks like all responders think all three of those states have more people than they do, even when no one response was biased that way.
With these two together, this survey makes the average (statistical mean) American look much less informed than the typical (statistical median) American.
Because at the end of it, they have a graph, with medians instead of means… and the same general pattern of significantly overestimating the size of minority groups and significantly underestimating majority groups is present.
Yeah, its less ‘strong’ overall when you look at medians instead of means, and some groups get pretty close to being estimated more in line with reality, but the same general pattern is still present.
So you could instead take this median based graph and order it by… gap between median estimated % and actual, and then do the same with averages, and compare and contrast that.
Like uh, heres a case study:
Mean estimated household income over $25k: 62%
Median household estimated income over $25k: 65%
Actual household income over $25k: 82%
Not too much difference, just 3%, between the mean and the median there, compared to:
Mean estimated household income over $1M: 20%
Median estimated household income over $1M: 10%
Actual household income over $1M: 0% *
Where 0% just means it is less than 0.5%, (or however they are doing rounding to nearest whole %) not literally 0.000000%.
So thats a more susbstantial estimated mean - median difference, 10%.
Before blaming me for not reading the article, maybe read my whole comment? I listed 2 effects. Yes, mean versus median is one. The other is a cognitive bias related to how humans estimate percentages.
Medians showing the same effect but reduced is exactly what I would expect when you account for one of the two phenomena but not the other.
You still clearly did not read the whole post, otherwise you would not have used the mean-median difference as an example.
That’s the more charitable intetpretation, by the way, the alternative is you did read it, knew it featured the median data points, and then acted as if it didn’t, ie, intentionally misrepresented.
Also, what you’re just asserting that round number bias is essentially uncontrolled for here? But you present 0 evidence of this.
That could be the case, but you have presented absolutely no evidence that that is the case.
As an econometrician, have you not heard of Kahneman and Tversky? That boggles my mind. Look up round number bias, and regression toward 50%, and maybe read Judgment Under Uncertainty: Heuristics and Biases.
This is mostly an example of a kind of survey bias I’m having trouble finding the name for plus a counterintuitive effect of averaging.
For the bias, when people are asked to estimate a percentage, they tend to estimate by large fractions of the whole, like by quarters, fifths, or tenths. This means you’ll see survey estimates closer to 50% than the real value, with the effect more pronounced for real values closer to 100% or 0%.
For the averaging phenomenon, when looking at the averaged responses across all questions of a survey, you can quite easily get a collection that wouldn’t make sense as a set of responses for “the average” (that is, the typical) person. You can have 3 different responders who each think California, Texas, or Florida has more people than they actually do, and then when you average those responses it looks like all responders think all three of those states have more people than they do, even when no one response was biased that way.
With these two together, this survey makes the average (statistical mean) American look much less informed than the typical (statistical median) American.
… You didn’t read the post.
Because at the end of it, they have a graph, with medians instead of means… and the same general pattern of significantly overestimating the size of minority groups and significantly underestimating majority groups is present.
Yeah, its less ‘strong’ overall when you look at medians instead of means, and some groups get pretty close to being estimated more in line with reality, but the same general pattern is still present.
So you could instead take this median based graph and order it by… gap between median estimated % and actual, and then do the same with averages, and compare and contrast that.
Like uh, heres a case study:
Mean estimated household income over $25k: 62%
Median household estimated income over $25k: 65%
Actual household income over $25k: 82%
Not too much difference, just 3%, between the mean and the median there, compared to:
Mean estimated household income over $1M: 20%
Median estimated household income over $1M: 10%
Actual household income over $1M: 0% *
So thats a more susbstantial estimated mean - median difference, 10%.
Before blaming me for not reading the article, maybe read my whole comment? I listed 2 effects. Yes, mean versus median is one. The other is a cognitive bias related to how humans estimate percentages.
Medians showing the same effect but reduced is exactly what I would expect when you account for one of the two phenomena but not the other.
You still clearly did not read the whole post, otherwise you would not have used the mean-median difference as an example.
That’s the more charitable intetpretation, by the way, the alternative is you did read it, knew it featured the median data points, and then acted as if it didn’t, ie, intentionally misrepresented.
Also, what you’re just asserting that round number bias is essentially uncontrolled for here? But you present 0 evidence of this.
That could be the case, but you have presented absolutely no evidence that that is the case.
These are fairly big survey/pollster people.
They probably know to account for that.
https://en.wikipedia.org/wiki/YouGov
We don’t know for certain though, because they do not extensively detail their methodologies in their post.
… but its… misleading to present the entire thing as being juat an artefact of not accounting for round number bias, or some other bias.
I’m an econometrician, it bugs me when people criticize statistics in invalid ways.
If you wanna wade into lies, damned lies, and statistics… arm yourself well, or, be a bit more humble.
As an econometrician, have you not heard of Kahneman and Tversky? That boggles my mind. Look up round number bias, and regression toward 50%, and maybe read Judgment Under Uncertainty: Heuristics and Biases.