A few days ago I ran the reader survey for 2024. You can take it here, but I probably won’t update the results again. In this post, I will look at the results.
In terms of the survey, there were 552 submissions, of which 238 clicked through to the end. Most questions were skippable so results for individual questions are based on fewer readers (you guys are subjects), but should still be decent. In comparison, the prior 2022 survey (I forgot 2023 apparently) had 318 submissions and 125 completions. So in terms of growth, an increase of about 90%. Pretty good!
Among submissions, 5% were paid subscribers. Of all the Substack subscribers, 1.3% are paid, so paid subscribers were unsurprisingly more likely to participate in the survey. Again, some 25 people pinky promised would sign up for paid, and I think 2 did, so less than 10% honesty.
In terms of when people started reading, it’s mostly recent, with a solid growth pattern. One person apparently has been reading since 2010 which is impressive. I wonder who that is. In 2010 I was completely unknown and had published no original science at all, and this was solely a philosophy blog. How things have changed!
I downloaded the subscriber data from Substack, and made this plot. There is a steady growth over time, and a slight upwards trend in the rate of growth. Right now, about 13 people sign up per day.
Using some simple models to forecast this growth, we can predict that by the end of the year, there should be about 10k Substack subscribers, a nice round number. Honestly, I am very honored by this growth. It is much more than I expected when I started writing on this platform in October 2021.
One thing is keeping up the writing and seeing growth in readership, but what about quality? We know from studies of music, movies, books etc. that creators tend to fall off over time and I’m turning 35 in a few days. So I asked people what they thought about the quality over time. The results are quite positive. Most people say there’s either no change or that they don’t know, presumably because they haven’t been reading long enough to say, but of those who say there’s a change, it’s a large majority in favor of the quality getting better. Only 3 people think it got worse, and those numbers are possibly just misclicks (or people very interested in philosophy).
I have occasionally asked people what they think I should be researching or writing about, and I came up with this ranking system. The results are pretty consistent over time (I posted it before on Twitter) in that most prefer the bread and butter to stay as it is: group differences research that few other people will study. There’s also solid interest in psychology of politics. While this topic is commonly studied by academics, most of this research is basically just trying to prove that conservatives are bad in some way or another (dumb, racist, closed minded). There’s relatively little interest in other differences, such as left-wingers being higher in mental illness, or libertarians being quite distinct personality wise.
Politics
We can also wonder about the politics of the readers themselves, and it looks like this. Most people are in favor of free markets. There’s a lot of variation around the degree of authoritarianism vs. libertarianism. There’s very few people from the extremes on the left, which is not too surprising.
In terms of labeling oneself, I gave people a list of such political labels and asked them which they would use if God forced them to be honest. No label is really dominant, it’s a mix of libertarian-likes (classical liberal, libertarian), nationalists (ethnonationalist, nationalist), and conservatives. But there’s also a good amount of people with other views, even a few (3) communists and (8) socialists.
Here I have computed the latent correlations between all the political labels. This is based on the assumption that each of these questions reflect continuously distributed variables that have been artificially dichotomized into yes/no (true/false) values. If one grants this assumption, one can compute what the correlation would have been had they been measured on a continuous scale instead of as true/false values.
We see that labels with conceptually related meanings do indeed cluster, which is to say, people who called themselves e.g. center-right also tend to call themselves YIMBY’s, globalists with correlations of about 0.4 to 0.6. Because some choices were very rarely chosen, the model has done some smoothing without which correlations cannot be computed. This can result in bizarre results, for instance, communist appears to be correlated with being a moderate! Maybe these are moderate communists. After all, they are reading this blog. So don’t get too excited about correlations with the rarely chosen labels. (Another example: feminist and anti-feminist correlate at -0.03, which is odd, as it should be more like -1, but there are only 6 feminists).
In terms of people’s own political changes, most people have moved to the right. I guess this is a response to the general increase in Wokeness and general social leftism.
Fertility, sex
I often write about fertility problems and patterns (eugenics, dysgenics), so naturally it makes sense to ask how many children readers actually have, so far at least. There’s a few people who say they have 1000, presumably fake values, but if we reduce the max value to 20, it looks like this. Many readers are quite young, so the average is not expected to be that high.
The median age is 35. So if we plot fertility by age:
So by the 40s, people have about 2 children on average. I am thus happy to report that we actually reach replacement fertility. Readers of the blog can sustain their numbers. (I removed the person who said they have 1000 children.) By the way, the readership is 94% male. We don’t have enough people to plot the fertility by sex or much of anything else, so I am going to keep using the complete sample and treat it mostly as being male. There were 2 transsexuals who took the survey, but fortunately both of them liked that.
We could also ask how many children people want to have. It turns our some people want to be Genghis Khan with 1000s of children and presumably a big harem. If we ignore these and focus on the normal range, the mean number is 3.5. Since readers appear to be getting about 2 by the end of their reproductive life, people are managing to have too few compared to their own wishes. This is actually a general thing, and is found across most countries.
Some people claimed to have had a lot of sex partners. I believe them (mostly), I know a few guys who are above 100, and I heard this is not so rare in the gay community. Still, there’s a big cluster at 0, which could mean incels, but it can also just mean young people.
Plotting claimed sex partners by age reveals an obvious age effect but the outliers are distorting the results. So let’s try modeling the median.
Here it seems we get more sensible results. The median (50th centile) person rises to about 10 life-time sex partners by their 50s.
About 93% were heterosexuals. Survey takers are slightly more heterosexual than my Twitter followers for some reason:
Race, ethnicity
As a new thing, I tried having people estimate their own ancestry. I gave them a list of options and asked them which % of their ancestry is from each. In theory, these numbers should sum to 100%. There didn’t appear to be a way force this in the survey software, but we can normalize the data so that each person’s ancestry sums to 100%. If we do, it looks like this:
Mainly readers are Europeans, with the biggest ancestries being those common in the USA, and an extra dose of Scandinavian. If we add up all the core Europeans, we get 85%, and if we add in the Jews we get 90%. In other words, this is a very European blog, as of course one would expect given the usual coffee salon demographics that Anatoly Karlin has written about.
It’s difficult to plot each individual’s ancestry since there are 19 ancestries and thus 19 colors if we tried a stacked bar plot. However, we can do this “stick” plot:
So we see that there’s quite a few purebreds of the different groups.
I asked people about their citizenship(s), but this data is somewhat of a mess to analyze. So here’s an easier solution based on the IP location of survey respondents. Most readers are from the big English-speaking countries, but the Scandinavians are overrepresented given their tiny population sizes. We can also confirm my usual observation that Norway is not doing very well in terms of dissidents. It has the same population size as Denmark (about 5 million natives), but has only 3 subscribers, while Denmark has 14. It’s not just a Danish-thing, Sweden has 26 and the population size is about twice that of Denmark and Norway (about 9M natives). We can thus also calculate a kind of dissident readership score by computing the per capita (per million in this case) readership rate:
Using a direct correction for population size results in very large standard errors for small populations, so it is not too surprising that a tiny country like Luxembourg managed to beat Denmark and Sweden. A different method is the log-count regression method, which is the residuals from this model: log(1 + readers) ~ log(population_size). This method is not intuitive but it does work (Noah Carl introduced me to this idea). If we use this method, then we get this result:
Since this method really just consists of trying to predict the log(reader count) by log(population), the resulting residuals of the model are telling us which countries have values that deviate from this expectation. Since small countries have small counts too, they don’t have very large deviations. For instance, Denmark has 14 readers, so the log(reader+1) value is 2.7. The population size predicted value is 0.40 so the value only deviates by about 2.3. For the USA, there are 256 readers, and the log value is 5.5, and the model predicts 0.77, so the deviation (residual) is about 4.8. From a Bayesian perspective, what this is doing is adding shrinkage so the values for small populations are moved towards 0. There are more sophisticated methods one could use based on count data modeling, I am not super familiar with this area of statistics.
Anyway, what the results show is that taking population size into account, Europeans read my blog a lot more than would be expected by population size. If you look at the other countries with low scores, you can also see that intelligence cannot explain this variation that well because countries with few northern or Germanic Europeans don’t read it much even if they have high intelligence (e.g. South Korea, Poland). This pattern is part of the WEIRD personality profile that Joseph Henrich, Jayman, HBD Chick, Peter Frost and others have been writing about.
Religion, education
In terms of religion, a majority are atheists (I consider agnostics to be atheists). This is somewhat impressive considering the fertility rate we see since atheists are usually quite low fertility. One might in this sense consider atheism a maladaptive human trait. I say this as an atheist.
In terms of education, the readers are very well educated. 14% have a PhD (US general population is about 1%), and there’s another 5% with long medical or law degrees. In fact, over 75% have university degrees. About 45% of readers are more educated than me.
Most readers were however just science consumers, not scientists themselves. But still, 26% were scientists and another 17% had degrees in it without using it.
Ratings
I asked people to rate 54 people, groups, or things.
In my eagerness to include all the interesting things, I neglected to think about how this might be plotted. After tinkering with it, this is maybe the best I could do. If you have very precise eyes, you can maybe work out that people who like Muslims (variable 5) also like the New York Times (9), and dislike Hitler (24). Instead we could factor analyze the ratings into 2 dimensions.
What this shows is that if you ask our friend the algorithm to find 2 hidden dimensions in the data which explain why the people who rate e.g. Hitler highly also rate Putin highly, etc. then it comes up with this pattern. This gives us a kind of position of each topic or person. Most of the dictators are in the bottom (Lee Kuan Yew is in the top center). Bottom left has the usual outgroups and their supporters (Muslims, leftist media, communist academics), top left has the intellectual Jews and their supporters (‘intellectual dark web’). Top middle has the HBD-themed topics, except for Joseph Bronski.
Alternative, we could ask how much people in general like the different targets:
Somewhat embarrassingly, I came out at top ahead of people’s ratings of their own parents, and of themselves. Substack also enjoys a really good rating. On the other hand, the most disliked targets were Mao, Stalin, journalists, Biden, and Hitler. In general, dictators were unpopular, as Putin and Xi Jinping also have poor ratings, Pinochet has about a neutral rating (-0.2, scale is from -2 to 2), and only Lee Kuan Yet has a positive rating (0.8). The results are definitely amusing.
Psychology
I asked readers to fill in their SAT scores, and results from other tests. The data are a bit of a mess since some people don’t follow instructions, but if we remove the bad data, the correlations are:
Most correlations look sensible. I have suppressed correlations based on fewer than 15 observations (hole in plot = sample size less than 15). Still, many correlations are based on small, and a very selective sample, so not all correlations are what we would expect, e.g. Mensa matrix vs. English vocab correlates at 0.09 whereas in a real dataset of native speakers, this correlation would be about .60. In fact, not too much can be made of this correlation matrix with the dataset as it is. Maybe in 2 years when the sample size has doubled again, we can find something more informative.
The mean IQ here is about 130 (127 Mensa test, 130 Wechsler). We could look up the norms to convert raw score on English vocabulary test, but 38 out of 45 must give a relatively high score. Likewise, the average SAT score was about 705. According to my conversion table, this corresponds to about the 97.7th centile and an IQ expectation of about 124. Personality wise, it’s medium-high openness, average conscientiousness, medium-low extroversion, medium-low agreeableness, medium-low neuroticism.
Other stuff
There has been some discussion about what the term human biodiversity should refer to. I agree that it’s commonly used to mean race and ethnic differences, but I think it should be properly taken to refer to all such biological differences, including between individuals. Readers did not generally agree with this usage, however.
Most people continue to use Python, but there’s a big spread in languages. For the record, coding is for nerds, R is master race, and Python is also good.
That’s roughly everything worth looking at. You can see more automatically generated summaries of the data here. It was a fun survey this year. Post ideas for next year in the comments.