Educational attainment and IQ over time
College graduates in 1940 were as smart as PhDs in 2018.
Summary:
Not Acceptable:
Acceptable:
A twitter thread has gone viral, claiming that the quality of the education system has decreased substantially in the last 80 years. The poster elaborates on how the standards and quality have declined throughout the years and that education is no longer as valuable as it used to be. They blame this on the GI bill, which introduced hordes of uneducated and unintelligent veterans into the college system and centralized the American education system.
Overall, I find little evidence that changing educational standards have impeded the scholastic achievement scores of children. First of all, there has been no decrease in scores on crystalized measures of ability (e.g. vocabulary, mathematics) ever since the 1950s (Flynn, 2007) or scores on measures of mathematics and reading from the NAEP (CATO, 2021).
Critics of modern education may say that the fact that scholastic achievement is not increasing as quickly as scores on other IQ tests is evidence that children are not being educated as well as they used to be. This rests on the erroneous assumption that scores on these tests are invariant over time - this is not the case as Flynn effects are not measurement invariant (Basement Inhabitant, 2023) and that Flynn effect increases are not on g (Nijenhuis, 2007).
Reviews of the literature surrounding the relationship between educational attainment and intelligence suggests that educational attainment improves scores on scholastically related tests, but not other tests (Hu, 2022). One study uses Swedish military test recruits as subjects, attempting to find the relationship between age and school days with the test scores. Independent of age, school days increased scores on synonyms and technical comprehension, while age itself was associated with higher scores on spatial and logic subtests independent of school days (Carlsson, 2015). This shows that the skills taught at school do not transfer to other cognitive abilities.
Jensen also notes this observation in his book Educability and Group Differences:
An interesting difference between scholastic achievement scores and intelligence test scores (including vocabulary) is that the latter go on increasing steadily throughout the summer months while the children are not in school, while there is an actual loss in achievement test scores from the beginning to the end of the summer. Much of the most recently learned material prior to the summer vacation has not been sufficiently rehearsed to become consolidated. The loss is greatest for those school subjects that depend least upon general intelligence (i.e., the consolidation factor) and depend most upon sheer learning and memory, such as spelling, punctuation, grammar, and mechanical or computational arithmetic and number facts, as contrasted with reading comprehension and arithmetic concepts (Beggs & Hieronymus, 1968).
As for whether intelligence itself causes differences in educational attainment between individuals, it does. This is because intelligence test scores measured at young ages are still able to predict educational attainment, and that associations between intelligence and success generally persist within siblings (Hegelund et al., 2019; Murray, 2002).
That isn’t to say I support the modern educational system in any way. The west has massively increased public spending on public education, with null to negligible returns to scholastic achievement. What I instead am attempting to articulate is that the modern educational system teaches skills just as well as the old one.
Others, such as Audacious Epigone, have found that the IQ scores within educational degree cohorts have been decreasing:
The problem with this analysis is that it uses the wordsum, a 10 item vocabulary test. I have evidence later in the article suggesting that this specific test may not be a great measurement of intelligence in modern cohorts (or even at all), and that analysis using the ASVAB increases the estimated scores of the most educated.
However, it is still possible to mathematically model the average IQs by educational degree by year using the following three variables:
How has the % of people in each educational attainment category changed over time
How has the correlation between education and IQ changed over time?
How has the average IQ of an American changed over time?
How has the % of people in each educational attainment category increased over time?
from here.
The correlation between educational attainment and IQ over time
For now, I have found two sources who have revised the relationship between IQ and educational attainment over time: Strenze and Zach Goldberg.
Here is Strenze’s table of moderators in his meta-analysis of success and IQ:
While the correlation between IQ and educational attainment appears to be decreasing, results from the regression table suggest this is not the case:
On the other hand, Zach Goldberg’s analysis of the GSS suggests that the correlation between educational attainment and wordsum score is steadily decreasing, and that the average score on the wordsum is decreasing as well:
It’s possible that this is because the validity of the wordsum itself is declining; lots of the words on the test are outdated and not used by the younger generations. I don’t think you will ever see the words ‘keyed’ or ‘transmaxxing’ in a vocabulary test - and hopefully they never will.
By observing whether the correlation between income and wordsum changes as well, you can test whether this phenomenon is unique to just educational attainment, or if it generalizes to other variables as well.
I was able to replicate his findings for both education and income, and show that the correlation is decreasing for income too:
Correlation between year and relationship with educational attainment: r = -.8, p < .001
Correlation between year and relationship with income: r = -.55, p < .01 (-.72 within men).
While the decrease in the correlation between wordsum and income doesn’t look as convincing, the relationship is much more drastic within men:
Looking back at the Strenze meta-analysis, the correlation between income and IQ does not appear change by year. Even so, I checked both the NLSY 79 and 97 and found that the validity of cognitive tests decreased by about .05 across the board between the two cohorts.
I found that the four differences in correlations I tested did reach statistical significance by any reasonable threshold:
The decreases in correlation observed using higher quality tests are lower (decrease of .05 vs decrease of .1) than the ones you see with the wordsum - leading me to believe that this test is not a good measure of cognitive ability in modern cohorts.
edit: Emil notes in the comments that the part of the decrease is due to more people not working - I assume this causes range of restriction of a measure by nature. Looking through the codebook, this appears to be true, given that a greater proportion of the NLSY97 is not working than the 79 sample (proof in Appendix).
How has the average IQ of an American citizen changed over time?
As many who read this blog know, the Flynn Effect describes a rise in IQ scores of about 2-3 points every decade (Trahan, 2014). The debate about these increases concerns what proportion of this effect is due to measurement variance (the test measuring different constructs at different times) and which proportion is due to true increases. This is a topic that is too complex to cover within a subsection of an article and is more appropriate to publish in a small book.
Regardless, my position on the Flynn Effect is that most of the observed increases are due to measurement variance. Beyond that, there was a small increase of about 5-10 points due to changes in nutritional quality and health in the first half of the 20th Century. After this, there is a small genetic decrease due to dysgenic fertility that is masked by measurement variance across different time cohorts.
These changes in intelligence are too negligible to take into account, especially when my data on educational attainment only starts in 1940. Because of this, I will be assuming in my analysis that the American IQ has stayed stagnant at 100 ever since 1940. For those who would like to learn more about the Flynn Effect:
Violation of measurement invariance across time in IQ tests: by Basement Inhabitant
Negative Jensen effect detected in Flynn Effect increases: by Nijenhuis and Flier
Schooling enhances IQ, not intelligence: by Meng Hu
Secular changes in development quotients: by Richard Lynn
The role of nutrition in secular changes in intelligence: by Richard Lynn
Brain size and intelligence (relevant to above study): by Emil Kirkegaard
Decrease in hue discrimination over time: by Woodley and Fernandes
Secular decreases in reaction time: by Woodley et al.
Criticism of above study by Scott Alexander and HBDchick
Response to the above criticisms: by Woodley et al. (I side with Scott and HBDchick)
Meta-analysis of dysgenics of intelligence: By Woodley et al.
The negative Flynn effect: By Dutton, Lynn, and Van der Linden
Flynn effect and its reversal are both environmentally caused: By Bratsberg
Simulation Results
Latent educational attainment (LEA) can be modeled as a normally distributed variable that is labeled in the real world as “dropout”, “high school graduate”, or “college graduate”, depending on how the individual ranks relative to his peers. Given that LEA correlates with IQ, I can then calculate the average IQ of each category. The correlation between IQ and education was fixed at .5 for every year due to contradictory evidence regarding whether the correlation has declined or not. The IQ of American citizens was fixed at 100 for similar reasons.
Currently, I only have longitudinal data regarding high school and college completion rates, so I can only simulate the results of 3 categories. Based on the NLSY97 data, the correlation between IQ and education decreases to .47 when you use 3 categories instead of 6, which must be taken into account
Here is the code I used to simulate the average IQs by education level for 2011:
To check how accurately these correspond on to real data, the NLSY97 was consulted and the same results were found. This means that any assumptions I am making within these simulations are not causing any discrepancies between the simulated results and the real data.
The educational attainment time series data was broken down by sex, so I took the liberty of breaking the results by sex. Most modern test IQ test batteries yield a sex difference of 4 points in adults; for reasons I will not reveal I assumed the difference was 3.3 points. When the simulations for every single year were executed, these were the results:
Note that these don’t imply the average IQ is decreasing (it was purposely held constant at 100); That would be a classic case of the Simpson’s fallacy.
I used the same simulation technique to calculate IQ scores by degree using more detailed data from 2018, and came to this conclusion:
Using educational attainment data from 2021 broken down by sex:
As for why there is such a pronounced sex difference - this is simply because women are more educated than men, but they score lower on IQ tests. Because of this, the difference you will see within educational attainment categories will be larger than the difference you see within the general population. These larger differences also replicate in the NLSY97, even though the female advantage in educational attainment is less pronounced in these years.
Appendix
Rpubs of the 5 R scripts (poorly formatted, as always):
Results of WORDSUM over time, within men:
Codebook card for 2011 income, NLSY97:
Codebook card for 1990 income, NLSY79:
Apparently some of the decline in the income ~ g, between NLSY 79 and 97 is due to a larger share of nonworking people.