{"id":15090,"date":"2026-03-27T11:53:02","date_gmt":"2026-03-27T10:53:02","guid":{"rendered":"https:\/\/emilkirkegaard.dk\/en\/?p=15090"},"modified":"2026-03-27T12:35:05","modified_gmt":"2026-03-27T11:35:05","slug":"dysgenics-for-whom","status":"publish","type":"post","link":"https:\/\/emilkirkegaard.dk\/en\/2026\/03\/dysgenics-for-whom\/","title":{"rendered":"Dysgenics for whom?"},"content":{"rendered":"

Some time ago, we published studies showing that the negative correlation between intelligence and fertility varies by other factors. Specifically, we found that it varies by religiousness and conservatism, in that leftists and unreligious people show stronger dysgenic effects<\/a>. These results were based on reasonably large samples. The largest was the cumulative GSS dataset with decades of representative Americans, resulting in tens of thousands of subjects, and still a large number of we subset to people with completed fertility (age 50+). Second, a replication in the Vietnam Experience Study, which has 4500 American former soldiers measured at age 38. This is somewhat problematic because men aren’t done reproducing at that age, also not in 1985 which is when the data is from.<\/p>\n

So I decided I wanted to dig into these findings and conduct a large-scale comprehensive replication. I surveyed Americans on Prolific and only targeted those with completed fertility in ages 50-69. I wanted to try out the Multifactor General Knowledge Test on another sample. It has an odd response format, where it shows you 10 options and says “Pick the 5 things that are X and don’t pick any things that aren’t X”. You get 1 point for each correct thing you picked and 1 point for each incorrect thing you didn’t pick. There are 32 questions, and thus one can get 320 points at most. We shall see that this format presents severe psychometric issues that I didn’t previously see, and Seb Jensen didn’t note them in our paper on this test<\/a>. In addition, I put together a few questions to measure overall religiousness, a standard US-style left-right scale, and a checklist for mental illness diagnoses. Thus, putting all of this together, we should be in a decent position to find those pesky effects. They are hard to find because the intelligence fertility correlation is weak to begin with, about 0.10, and we are looking for relatively week interactions with that, say, that changes it to 0.00 or 0.20. This require a lot of statistical precision.<\/p>\n

This post is basically the preview of the academic write-up of this article, since I just finished the analyses after waiting some months for the data collection to finish.<\/p>\n

This was the most annoying part. Naturally, you would think that since selecting the right things (say, American poets) out of a list of 10 options requires knowledge of such matters, and so does not picking the wrong options, these correct decisions should show positive correlations. But they don’t actually. They show negative correlations, in apparent violation of the positive manifold. Why is that? It’s because of guessing. Some people are more inclined to guess than others, and those people get\u00a0both<\/em> more correct selections and more incorrect ones, thus creating negative correlations between them. We can see that by looking first at the pass rates by item:<\/p>\n

\"\"<\/a><\/p>\n

The correct selections have higher pass rates, so people are better at picking the right options than they are at not picking the wrong ones. Their loadings distributions if we analyze with a single factor for all items (2PL model):<\/p>\n

\"\"<\/a><\/p>\n

You see the problem. And this is not because they are not scored correctly, they truly have reverse of expected loadings because of this format. We can also get the same pattern at the aggregate level by summing the positive and negative scores by item (thus 32*2 mini-sums):<\/p>\n

\"\"<\/a><\/p>\n

Too small to read, they are in pairs, so 1st variable is poets for the positives (select them) and 2nd variable is poets negative. This pair actually has a positive correlation, but in general, the pairs don’t have positive correlations (blue color). For good measure, I went back to the old dataset and made the same plot to verify:<\/p>\n

\"\"<\/a><\/p>\n

It looks materially the same, so this pattern is not caused by Prolific users, but is also seen in these online test takers from openpsychometrics.org<\/a>.<\/p>\n

There’s a number of ways we could then score these data. Fortunately, we don’t have to guess from theory which one is best because we can simply check the correlations of the various scores with our other variables. The one that results in the strongest correlations is the best method empirically. We can use religiousness as our criterion given that it has a moderate negative correlation that is very well established. The sum score obtains a correlation of -0.32, higher than the other methods.<\/p>\n

The left-right and religiousness scales didn’t present any notable problems, just standard single factor good solutions. For the mental health items, since they are binary, we can plot their latent correlations (tetrachoric):<\/p>\n

\"\"<\/a><\/p>\n

Like cognitive data, the mental health checklist shows an mostly perfect positive manifold. The question is then whether a fancy IRT score (a weighted sum) can beat just taking a simple sum. We can compare the results of the scores:<\/p>\n

\"\"<\/a><\/p>\n

They correlate at 0.96 but the linear model is inappropriate. Possibly this matters for some analyses.<\/p>\n

With this stuff done, we can now plot the correlation matrix across our variables:<\/p>\n

\"\"<\/a><\/p>\n

Most things look normal, e.g., worse mental health correlates with female and leftism, religiousness correlates with fertility and negatively with leftism. The most surprising is the ~0 correlation between mental health and IQ, normally it’s negative at about -0.20<\/a>. It may be because of this particular measure (general knowledge means a lot of time spent reading instead of socializing) or because of the self-selection of sampling older people filling out surveys on the internet for 8 USD\/hour.<\/p>\n

Fertility, this was just a question about how many biological children people have. We replicated the usual mysterious male-female difference:<\/p>\n

\"\"<\/a><\/p>\n

It may be due to self-selection bias with the sampling. Notice also that the data do not fit the Poisson distribution exactly, there are too many 0’s. This has some implications for modeling. For predicting fertility we have some options:<\/p>\n