Before we start, Turkheimer and I disagree about the utility of IQ and the existence of general intelligence. Whereas I believe IQ is dead, and general intellig

Before we start, Turkheimer and I disagree about the utility of IQ and the existence of general intelligence. Whereas I believe IQ is dead, and general intelligence is a made up construct, Turkheimer still believes general intelligence exists and IQ is a good way of measuring it. What Turkheimer has done however is show us a way to use IQ while not going into the race science of it. As such I think he has done a terrific job, even though I disagree with him philosophically. Now just to clarify, G + E is a different model from G + E + G×E. Many twin studies in the past century were done on a pretty homogeneous demographic, across a very slim slice of SES (mostly inside high SES; as you correctly pointed out). This skewed the models such that it completely missed the G×E part, this is, if you don’t measure across different SES, you’ll miss how SES interacts with genotypes. What Turkheimer did was actually measure heritability inside low SES, and his results were that in a G + E model, the proportions were flipped. That is, instead of G=60% and E=40%, he and his team found E=60%. When your get results that flip in this way, it is a pretty clear indicator that you have interactions between your main effects (G and E), so Turkheimer fitted a model which includes the G×E interaction[1] (which all researchers with some level of statistical literacy would do) and found that almost all of the variance previously attributed to genes, would perish into the interaction effects. You can read the study here. The methodology is not that far off from what you are proposing (just without the controlled interventions; which is bound to cause some biases anyway). https://www.researchgate.net/publication/8997472_Socioeconomic_Status_Modifies_Heritability_of_IQ_in_Young_Children https://www.researchgate.net/publication/8997472_Socioeconom... --- 1: Note the models are a bit more complicated than G+E. Usually they use an ACE model, with Genotype (A), shared environment (C), and non-shared environment (E; also known as gloomy prospect), and than model A + C + E. In Turkheimer, et. al. (2003) the model was s*SES + A*(a + a'*SES) + C*(c + c'*SES) + E*(e + e'*SES) where a,c,e are the main effects of A, C, E; s the main effect of SES; and a', c', e' are interactions of a,c,e with respects to SES.