In 1300 the homicide rate was about 50 per 100,000 people, or 0.5 per thousand. Homicide must have caused on the order of 1 to 2 percent of all deaths and a much higher proportion of deaths of young adult males. Our assumption of a Normal distribution of the underlying trait immediately implies that the threshold was 3.3 standard deviations greater than the mean (from any table of the Normal distribution). Natural selection, social selection we would say in this case, disfavors homicide and the distribution is shifted so that the homicide threshold is surpassed by only 1 in 100,000 people rather than 1 in 2,000. By the year 2000 the homicide threshold is at 4.3 standard deviation from the population mean. In other words selection has moved the distribution 1 standard deviation in 700 years or 28 generations. …
In the present case we need a response of 1/28 of a standard deviation per generation. Assuming an additive heritability of 0.5 (the true value is probably 0.8 or so from literature on the heritability of aggressive behavior in children) the selective differential must be about 1/14 or .07 standard deviations per generation. In terms of IQ this would correspond to a one point IQ advantage of parents over the population average and in terms of stature parents with a mean stature 0.2 inches greater than the population average. This would occur if the most homicidal 1.5% of the population were to fail to reproduce each generation.
If homicides were responsible for one to two per cent of all deaths around 1300, it seems, on face value, unlikely that the most homicidal 1.5% would fail to reproduce. If the effect was direct, that would require most murderers to reach an evolutionary dead-end. But as Harpending notes in the comments to his post:
[W]e think (most of us) that homicide is an indicator of something else—proneness to violence, high time preference, fast life history, and so on. We cannot treat homicide as a discrete real trait. It is instead a shadow on the wall of the cave.
If those who had lower rates of time preference and who were less risk seeking gained a fitness advantage from success in apparently unrelated domains such as commerce, the effects would almost certainly play out in the context of violent behaviour.
Harpending’s post also reminds me that quantitative genetics is a severely underused tool on questions such as these.