{"id":15112,"date":"2026-03-28T09:06:58","date_gmt":"2026-03-28T08:06:58","guid":{"rendered":"https:\/\/emilkirkegaard.dk\/en\/?p=15112"},"modified":"2026-03-28T09:06:58","modified_gmt":"2026-03-28T08:06:58","slug":"consequences-of-racial-composition-changes-in-usa-1970-2023","status":"publish","type":"post","link":"https:\/\/emilkirkegaard.dk\/en\/2026\/03\/consequences-of-racial-composition-changes-in-usa-1970-2023\/","title":{"rendered":"Consequences of racial composition changes in USA 1970-2023"},"content":{"rendered":"

Some years ago, I gave a good idea of a study to a friend, and he did it. But then he just never published his study anywhere, it’s just sitting there in a Google Docs file that no one can read. So I decided with the power of Claude to redo and expand on the study. The annoying part of doing most research is collecting the data files and dealing with formatting, finding the right variables etc. But Claude will do that quickly, then it’s just a matter of figuring out the right models, good plots. In other words, AI does the boring part and human gets to do only the fun parts.<\/p>\n

So everybody knows that race predicts crime, that is, members of some races are more likely to be criminals than others. These rank-orders are mostly not but entirely consistent across time and place, which I reviewed previously<\/a>. In the USA, Blacks (~80% African ancestry, ~20% European) have by far the highest crime rates with a homicide rate about 6-10x the White one, depending on time<\/a>:<\/p>\n

\"You<\/p>\n

Given this strong association at the individual level, the geography of crime in the USA is to some extent a map of where Blacks live:<\/p>\n

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

The other red spots are Amerindian places, mainly reservations, and these are very sparsely populated. This isn’t so obviously to people who haven’t been there, but it looks like this:<\/p>\n

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

Nevertheless, someone might say: Blacks just commit crime because they live in particular places that make them do crime. It has nothing much to do with the demographics, but the [local something\/magic dirt]. The standard solution to this problem is to run some cross-sectional regressions controlling for other supposed causes of crime, which could be inequality or poverty:<\/p>\n

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

The betas are standardized, so they are comparable in a way. We can see that demographics is a much better predictor of homicide rates than are purely socioeconomic factors (60% variance vs. 34%). Nevertheless, since the predictors are not necessarily uncorrelated, standardized betas can be misleading about unique contribution to the model’s explanatory power. For this purpose, we can compare the epsilon values:<\/p>\n

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

These are from variance decomposition into partial epsilon\u00b2 values, like the more standard partial eta\u00b2. The issue with eta\u00b2 is that it is not adjusted for overfitting, so epsilon\u00b2 is the adjusted (shrunk) version of eta\u00b2, and taking the square root puts them in correlation-like units. Thus, we see that Black% is about 4x more important than the next predictor in explaining between county differences in homicide rates.<\/p>\n

Many studies exist with this approach, and Black% always beats the other predictors. With regards to causality, on the hereditarian model, such models are wrongheaded because controlling for poverty indirectly controls for the human capital of the Blacks. Blackness (or Africanness), of course, in itself does not really cause anything, rather it is because the races have different heritable statistical distributions of psychological traits — intelligence, aggressiveness, sociability, and others — that cause these differences in poverty, education, crime rates and so on.<\/p>\n

Still, maybe the Black% effect is due to something else, something we didn’t measure. That can be taken into account using spatial regression models:<\/p>\n

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

The spatial variable captures anything unobserved that is similar for nearby counties (unobserved spatially autocorrelated confounders), and it has a quite strong effect. But it doesn’t really change the differences among the other variables.<\/p>\n

The final approach is to use what economists call panel data, and what everybody else calls a longitudinal fixed effects model. Instead of trying to measure confounders (OLS), or capture them via spatial structure (spatial model), we can exploit variation across time for the same units. In other words, does change in Black% predict change in homicide rate? Here’s results across approaches:<\/p>\n

\"\"<\/a>The outcome variable here is in its natural unit, the homicide rate per 100k, and Black% is in its natural unit, fraction of the population. This means that the slope can be interpreted as hypothetical change in homicide rate if a county was magically changed from 0% to 100% Black. We see that across the models, they all estimate a large positive value compared to the national average which is 5.7. Curiously, the more strict panel models find larger effects. There are two tows for the panel data because the data from before 1990 does not have proper Hispanic questions, so we can’t calculate non-Hispanic Black% non-Hispanic White% etc. This is also why we can’t use the panel data with Hispanics from before 1990 (I tried, and got very unstable results).<\/p>\n

For good measure, here are the results for the other socioeconomic variables using the cross-sectional with state fixed effects:<\/p>\n

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

For comparison purposes, the outcome variables have been converted to z scale. Reading the numbers thus tells you that compared to having 100% Whites, magically changing to 100% Black lowers median income by 2.8 standard deviations, or going from 100% Whites to 100% Asians decreases the poverty rate by 2.9 standard deviations. Numbers for Asian% tend to generate extreme results since there is very little variation in the data as most counties only have trace amounts of Asians. By the way, the strangely ~0 result for Black% and education is because of high White% rural poor counties with little education (Appalachia). Here’s the panel models for the same outcomes:<\/p>\n

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

Gone are the extreme Asian% values, mostly, and so is the Black% null effect on education, but we get the new abnormality that Hispanic% predicts lower<\/em> homicide rates. We know that actual national statistics that Hispanics commit homicide at about 2.5x the rate of Whites, so that value we would expect to recover (if we scaled the numbers correctly), at least in theory (assuming a purely compositional model). Nevertheless, of the 20 betas estimated using the panel data, 19 are in the expected direction. In other words, looking only at changes in demographics of states, one can predict their futures with some accuracy. I don’t know what’s up with the Hispanic finding, I guess they are moving into areas that are already falling in crime.<\/p>\n

In general, then, the panel data support the usual plotposting where causality is just asserted. If anything they produced larger<\/em> slopes than the simpler models. The effects are qualitative large, e.g. for the homicide rate:<\/p>\n

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

Going from 0% Black to 100% means a homicide rate change from 2.5 to 27, or about 12x. This is larger than the difference we know to be the ground truth, about 8x. I think this happens because of self-selection. The Blacks in heavily Black areas (inner city ghettos, deep south) are below average among Blacks, the good Blacks move to less Black areas, which thus inflates the slope. Additionally, the Blacks in less Black areas are genetically less African. The effects for the other outcomes are similarly large (e.g. 3.7 standard deviations higher poverty rate), but this aligns with the very large differences we are already familiar with between countries inhabited by different peoples.<\/p>\n

Demographics is destiny.<\/p>\n","protected":false},"excerpt":{"rendered":"

Some years ago, I gave a good idea of a study to a friend, and he did it. But then he just never published his study anywhere, it’s just sitting there in a Google Docs file that no one can read. So I decided with the power of Claude to redo and expand on the […]<\/p>\n","protected":false},"author":17,"featured_media":15116,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2351,3636,1937,1921],"tags":[2322,3833,2149,2100],"class_list":["post-15112","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-criminology","category-demographics","category-education-science","category-sociology","tag-counties","tag-fixed-effects","tag-states","tag-united-states","entry","has-media"],"_links":{"self":[{"href":"https:\/\/emilkirkegaard.dk\/en\/wp-json\/wp\/v2\/posts\/15112","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/emilkirkegaard.dk\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/emilkirkegaard.dk\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/emilkirkegaard.dk\/en\/wp-json\/wp\/v2\/users\/17"}],"replies":[{"embeddable":true,"href":"https:\/\/emilkirkegaard.dk\/en\/wp-json\/wp\/v2\/comments?post=15112"}],"version-history":[{"count":1,"href":"https:\/\/emilkirkegaard.dk\/en\/wp-json\/wp\/v2\/posts\/15112\/revisions"}],"predecessor-version":[{"id":15122,"href":"https:\/\/emilkirkegaard.dk\/en\/wp-json\/wp\/v2\/posts\/15112\/revisions\/15122"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/emilkirkegaard.dk\/en\/wp-json\/wp\/v2\/media\/15116"}],"wp:attachment":[{"href":"https:\/\/emilkirkegaard.dk\/en\/wp-json\/wp\/v2\/media?parent=15112"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/emilkirkegaard.dk\/en\/wp-json\/wp\/v2\/categories?post=15112"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/emilkirkegaard.dk\/en\/wp-json\/wp\/v2\/tags?post=15112"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}