4 ms·

I believe IQ scores are rising still for lower scores, but falling for the top 1-2%?

gwern 7y ago
The very most recent US Flynn effect study finds the opposite pattern, actually (falling for low end, rising for high end): https://www.gwern.net/docs/iq/2019-platt.pdf https://www.gwern.net/docs/iq/2019-platt.pdf Personally, I don't buy it. The Flynn effect is large enough, and the way the normal distribution works with tail effects, that if there were any Flynn effect in general (much less focused on the high end), that you should see dramatic effects like several-fold increases in the number of perfect SAT scores. You don't see anything remotely like that - even despite loads of countervailing factors like increases in SAT access and prep!
- SamReidHughes 7y ago
  Intelligence isn’t normally distributed though. E: And even if it was, and was capable of “shifting” left or right, which is a nonsensical notion about populations, a shift of a few IQ points wouldn’t have a dramatic many-fold increase in perfect scores.
  - gwern 7y ago
    You're right, intelligence, like anything else (eg height), is not exactly normally distributed: it has fatter tails than a normal distribution - so a shift in the mean would increase the density at a high threshold even more than you would calculate based on a normal.
    - SamReidHughes 7y ago
      No, a fat tail would do just the opposite. A fat tail is where more points are far from the threshold. But it's a moot point because an increase of 2 points among people in the 99th percentile implies little about those in the 99.99 percenteth quantile. Some description of a fat tail from 60-99 percentile doesn't say much about the top one percent. At a certain point, the dating market tops out.
      - pmart123 7y ago
        I believe it is the same case with height. https://www.johndcook.com/blog/2008/07/20/why-heights-are-not-normally-distributed/ https://www.johndcook.com/blog/2008/07/20/why-heights-are-no... It's close to a normal distribution one or even two deviations away from the mean, but real-life has way more samples appearing +5 deviations away from the mean. I haven't really looked into it ever, but this could be because of a positive skew versus a distribution with purely fat tails?
        SamReidHughes 7y ago
        The difference is, it's easy to make a long-boned freak or a blithering idiot. But there's no hormone (I know of) that makes you a genius. (Also, intelligence/aptitude/performance scales like IQ, SAT, or ASVAB scores are arbitrarily numbered, so I'm don't know what a "fat tail" means in that respect. Maybe there's a standard definition that isn't completely arbitrary.)
      - gwern 7y ago
        No, it would. Simulate it at a threshold like 1600 scores. The steeper increase in that part of the range yields the same phenomenon as normals. > (Also, intelligence/aptitude/performance scales like IQ, SAT, or ASVAB scores are arbitrarily numbered, so I'm don't know what a "fat tail" means in that respect. Maybe there's a standard definition that isn't completely arbitrary.) Er, you can look at plenty of cardinal-scaled things like vocabulary or digit span. Doesn't change the non-normality at the tails.

The very most recent US Flynn effect study finds the opposite pattern, actually (falling for low end, rising for high end): https://www.gwern.net/docs/iq/2019-platt.pdf https://www.gwern.net/docs/iq/2019-platt.pdf Personally, I don't buy it. The Flynn effect is large enough, and the way the normal distribution works with tail effects, that if there were any Flynn effect in general (much less focused on the high end), that you should see dramatic effects like several-fold increases in the number of perfect SAT scores. You don't see anything remotely like that - even despite loads of countervailing factors like increases in SAT access and prep!

SamReidHughes 7y ago
Intelligence isn’t normally distributed though. E: And even if it was, and was capable of “shifting” left or right, which is a nonsensical notion about populations, a shift of a few IQ points wouldn’t have a dramatic many-fold increase in perfect scores.
- gwern 7y ago
  You're right, intelligence, like anything else (eg height), is not exactly normally distributed: it has fatter tails than a normal distribution - so a shift in the mean would increase the density at a high threshold even more than you would calculate based on a normal.
  - SamReidHughes 7y ago
    No, a fat tail would do just the opposite. A fat tail is where more points are far from the threshold. But it's a moot point because an increase of 2 points among people in the 99th percentile implies little about those in the 99.99 percenteth quantile. Some description of a fat tail from 60-99 percentile doesn't say much about the top one percent. At a certain point, the dating market tops out.
    - pmart123 7y ago
      I believe it is the same case with height. https://www.johndcook.com/blog/2008/07/20/why-heights-are-not-normally-distributed/ https://www.johndcook.com/blog/2008/07/20/why-heights-are-no... It's close to a normal distribution one or even two deviations away from the mean, but real-life has way more samples appearing +5 deviations away from the mean. I haven't really looked into it ever, but this could be because of a positive skew versus a distribution with purely fat tails?
      - SamReidHughes 7y ago
        The difference is, it's easy to make a long-boned freak or a blithering idiot. But there's no hormone (I know of) that makes you a genius. (Also, intelligence/aptitude/performance scales like IQ, SAT, or ASVAB scores are arbitrarily numbered, so I'm don't know what a "fat tail" means in that respect. Maybe there's a standard definition that isn't completely arbitrary.)
    - gwern 7y ago
      No, it would. Simulate it at a threshold like 1600 scores. The steeper increase in that part of the range yields the same phenomenon as normals. > (Also, intelligence/aptitude/performance scales like IQ, SAT, or ASVAB scores are arbitrarily numbered, so I'm don't know what a "fat tail" means in that respect. Maybe there's a standard definition that isn't completely arbitrary.) Er, you can look at plenty of cardinal-scaled things like vocabulary or digit span. Doesn't change the non-normality at the tails.