IQ tests measure relative to a reference population. So we can speak of the IQ of a person aged 20 born in 1990 on a test where 100 is the average score at age 20 of the population of those born in 1960, etcetera.
Changes in average IQ has been something that has been spoken of in the literature since Flynn was young, back when IQ scores were rising.
So now we need to be really careful about bias in the sample groups.
To say IQ scores are dropping and we don't know why then means to me we don't know what variables were the key drivers of the first groups results so we therefore cannot possibly create a consistent second group to measure against the baseline.
I'm still stuck but I'm often dense and skeptical on these type of analysis.
This is why there are practices around statistics sampling and confidence intervals and whatnot. Of course we can’t exactly recreate, that’s impossible, but we can create samples in a way that the results are generalizable in both the baseline and the comparison. And the results are useful.
I have a friend who says that it’s impossible to know if we can’t sample 100% (“how can we know how it will work in 100 million if we only test in 1000?”) and it’s frustrating how just basic concepts of math and statistics make it hard for them to accept any research.
Another source is the number of clinical cases of mild mental retardation. This will mostly depend on how large the population subgroups with lowest IQ are, but if they are approximately fixed and there's no explosion in assortative mating it would be possible to calculate the true mean from that.
This assumes both classification criteria and testing rates remain constant over time. Both are poor assumptions.
It’s easy to see how such a system can be used to measure a decrease or increase of height over time, even though your normalizing your reporting and even though the markings are irregular.