The race differences in intelligence found in this dataset are smaller than those seen in representative samples (see Lynn, 2006). This is probably due to the use of self-selected samples. After adjustment for imperfect reliability, the gaps were still somewhat smaller than representative samples, e.g. the Black-White gap was 0.66 d versus about 1.00 d (IQ 15) typically found (Roth et al., 2001, Frisby & Beaujean, 2015, Fuerst et al 2021, Lasker et al., 2019, Kirkegaard et al., 2019). First, we relied on self-reports of gender and sexual orientation. These are both to some extent fallible as indicators of actual behaviour, e.g. some men who report being heterosexual have sex with other men (e.g. for pay), some men who report being women do so in order to enter female prisons and rape women (e.g. Howes, 15th January 1922). Future research would benefit from employing objective assessments of sexual orientation, such as by measuring levels of arousal in response to specific sexual stimuli. Second, our intelligence measure was crude with a reliability of only 0.62. This presents relatively few problems for our results because it is easy to adjust for imperfect reliability of the outcome variable in a regression model. Still, given the sampling of items, and people?s ability to skip items if they suspected they got them wrong, the measure is still suboptimal. Accordingly, it would be useful to re- conduct this study with a instrument that has higher reliability. Third, our dataset was large, but mostly limited to Western, English speaking populations. Because it was sampled from a dating site, it is also somewhat limited in age towards the younger side, and of course, is necessarily limited to people who spent their time on a dating site. There is however no particular reason to think these selection factors impacted our results much,
aside from the smaller than usual race/ethnic gaps observed. Nevertheless, any future study would benefit from employing a more clearly representative sample.