Milky Eggs

The tenured professor and the sea of cancellation: Gregory Clark on human intelligence

Gregory Clark is a tenured professor of economic history at UC Davis known for studying intergenerational mobility in England from the 17th through the 21st century. Interestingly, he more or less re-derived the additive inheritance of human talent (principally but not wholly composed of human intelligence) from empirical analysis of how social standing fluctuated throughout the centuries. These findings are well summarized in a recent podcast between him and the physicist Steve Hsu. I will highlight and expand upon some points of considerable interest from this podcast.

My primary aim will be to excerpt the most compelling parts of the (rather long) interview and provide some supplementary annotations, references, and commentary. As such, the content here may not be very interesting to those who are already experts in the genetics of human ability or familiar with Gregory Clark’s work. To all others, I hope that it may serve as a useful primer to some very fascinating lines of inquiry.

Highlights

An even briefer summary of the content below:

• Intergenerational status mobility can be estimated precisely over the last 3 centuries of English history
• Intergenerational status is consistently very immobile, with a correlation of r = 0.8 between generations persisting through the centuries
• A high degree of assortative mating on underlying status, to the tune of r > 0.65, is also consistent throughout the historical data, consistent with the degree of assortment on educational attainment polygenic scores observed in the UK Biobank
• Family environment is essentially irrelevant to long-term outcomes, with minimal effect sizes for family size, birth order, presence or absence of relatives, marginal years of education, or even the untimely death of a parent
• The “genetic hypothesis” of status transmission is consistent with modern work from other research groups studying both America and 20th century China

The remainder of the post remains highly recommended.

Calculation of intergenerational mobility

How might one derive valid results on intergenerational mobility that span the centuries? This is prima facie a quite challenging task as the nature of work, income, and professional status all vary greatly within even a single century alone. Moreover, we are really concerned not with a narrow measure like income or wealth alone, but a broader sense of status, which may certainly include income but also encompasses notions such as the reputation afforded by a given profession. Finally, there are complex issues of data quality, where a given annotation of someone’s professionーsay, bricklayerーmay in principle encompass quite a wide range of potential status outcomes (compare the bricklayer for the village outhouse versus one employed by the Crown).

The solution Clark uses is clever. In essence, he exploits the rich correlational structure of status-based matching (marriage) and transmission through family lines. For example, marriage choices are used to impute a rank-ordering of occupational status. Similarly, the structure of within-family status transmission can be exploited to adjust for error originating from ambiguity within broad professional categorizations.

I will not discuss the specific technical details here (for I am in truth not familiar with them). Nevertheless, the specific numerical results are quite interesting; for example, Clark and Cummins (2014), Intergenerational Wealth Mobility in England, 1858–2012: Surnames and Social Mobility claims the following:

This article uses a panel of 18,869 people with rare surnames whose wealth is observed at death in England and Wales 1858–2012 to measure the intergenerational elasticity of wealth over five generations. We show, using rare surnames to track families, that wealth is much more persistent than standard one generation estimates would suggest. There is still a significant correlation between the wealth of families five generations apart. We show that this finding can be reconciled with standard estimates of wealth mobility by positing an underlying first order Markov process of wealth inheritance with an intergenerational elasticity of 0.70–0.75 throughout the years 1858–2012.

What Clark is suggesting here is that when one naively looks at unadjusted measures of intergenerational mobility, the results actually enormously overstate the degree of mobility because the measures of status are insufficiently precise (which consequently deflates the intergenerational correlations); however, when various sources of error are adjusted for, we actually obtain very consistent intergenerational elasticities on the order of 0.7-0.8 across multiple centuries.

This is quite notable as, obviously, the world has liberalized considerably over the past three centuries! Naively, one would expect mobility to have increased dramatically. However, this does not seem to be the case. I cannot help but wonder if the “true” intergenerational status correlation is even higher than 0.8ーis it not plausible that Clark has only accounted for some of the measurement error, and that a perfect metric of social status would show even tighter transmission?

Human talent as an underlying factor

One might wonderーhow is it actually possible for the world to change so much, yet for intergenerational status transmission to remain so robust across the centuries? Recall that public education was almost unheard of in the year 1700!

A series of circumstantial observations suggests that there is some underlying additive genetic factor generating social status which is transmitted from parent to child and which plays a major role in the marriage matching process. While it is ultimately impossible to measure the IQ of someone from the year 1700 based on sparse historical records alone, these observations make the genetic talent hypothesis quite compelling.

For example, intergenerational correlations seem to decline exactly as genetic relatedness would imply:

Greg Clark: Right. I mean, what’s happening is that the correlation is declining as we move out on the family tree, but in a very predictable fashion, right? That always, as we move, as we move from second cousins to third cousins, that’s two steps away on the family tree, that correlation declines by a factor of 0.64.

[…] It doesn’t matter if you’ve ever met the other people, if you’ve ever had any connection to them, if you ever had any involvement with themーbasically, it shows this very regular structure that we can plot, on one axis people’s genetic relatedness, and then on the other, the underlying correlation between them. And it’ll just fall along the straight line.

We should not expect to see such a robust pattern manifest if immediate environment were the dominant factor in producing future outcomes. In a world where parental environment is paramount, one would naively expect moving from second to third cousins (all of whom have plausibly quite different parental environments) to correspond to a much lower decline in status correlation. Instead, the lion’s share of the correlational decline should be found when moving from immediate siblings to first cousins.

Another remarkable fact is that Clark’s work actually anticipated, before it was known, the degree of assortative mating on intelligence in modern England! That is to say, for Clark’s estimates of intergenerational mobility to arise from a standard, additive, polygenic trait, they imply that this trait must be a very important factor in mate selectionーmore important than was previously appreciated at the time Clark published his findings. And yet, some years later, large-scale studies on the UK Biobank exactly validated Clark’s implied prediction of assortative mating to the degree of r ~= 0.65 for intelligence:

Steve Hsu: Yes, so when I first saw some of your early work on just the rate of regression to the mean intergenerationally, and, you know, your results show that this was much slower than people would have expected. I myself thought, this can’t be right, because even if I assume a very high heritability for the underlying traits, surely the degree of assortative and mating is not high enough to actually give the results that you found empirically.

And so, I was just amazed when I saw this more recent paper by you and also looked at the genomics results from UK Biobank, which you just mentioned, which actually show that the level of genetic assortativeness is extremely high. It’s something like 0.65, the correlation between the polygenic score for educational attainment in married couples.

So, I was actually amazed that your empirical work implied this result many years before we were actually able to measure it in actual genomics.

Indeed, as Steve says, Robinson et al. (2017), Genetic evidence of assortative mating in humans finds:

We extend our analysis to the UK Biobank study (7,780 pairs), finding evidence of a correlation at trait-associated loci for waist-to-hip ratio (0.101, 0.041 SE), systolic blood pressure (0.138, 0.064 SE) and educational attainment (0.654, 0.014 SE). Our results imply that mate choice, combined with widespread pleiotropy among traits, affects the genomic architecture of traits in humans.

An even stronger degree of matching on intrinsic ability is obtained in Clark’s newest work, For Whom the Bell Curve Tolls: A Lineage of 400,000 English Individuals 1750-2020 shows Genetics Determines most Social Outcomes. Under simple assumptions about the correlational structure between the family members on both sides of a marriage, he calculates the degree of assortative mating on underlying, status-based characteristics as approximately 0.8:

Although it superficially appears that these correlations are declining in the modern era, that figure is somewhat suspect as the number of observations is much more limited than for prior ranges of time.

It is interesting to note that even back in the early 19th century, well before it was standard for women to attend university or work in high-skilled professions, the degree of matching on ability is already as high as r = 0.8! Still, from another perspective, it makes senseーromantic compatibility is very much a function of intrinsic characteristics which can be mutually ‘measured’ to rather high precision with even several hours’ worth of interaction, and it is only natural that even without the explicit aid of status markers like university degrees, humans find a way to assort on these vital characteristics.

The surprising irrelevance of environment

The primacy of genetic influence is further reinforced by several remarkable findings about the relative irrelevance of family environment on long-term success. Consider, for example, family size: one might naively assume that large family sizes diminish the number of resources available per child, thereby reducing prospects for success (especially in pre-modern England, where family planning and birth control were uncommon and so family size was effectively randomly distributed).

Clark explicitly addresses this question in For Whom the Bell Curve Tolls:

While larger family sizes appear to negatively affect educational outcomes, their effects on occupational status are substantially muted in comparison, consistent with a model where underlying traits ‘reassert’ themselves in adulthood and the importance of early-life environment fades into the background. (The negative effects on wealth are plausibly due to factors such as inheritance effects.)

The absolute levels of influence on family size on occupational status are themselves quite modest:

For families in the lineages of average or poor social status expanding family size from 1 to 12 reduced the adult occupational status of children by 7%. For rich families the estimated effect was greater, an 18% reduction. But in both cases the overwhelmingly strong predictor of social outcomes was the social status of the father, not the numbers of children in the family.

A separate working paper by Clark and Cummins, Family Matters? Do Relatives other than Parents Matter to Social Outcomes, England 1780-2016?, claims that in general, the presence of non-parental relatives does not seem to causally matter: i.e., the correlation between a relative’s status and a child’s status (beyond that which is attributable to parental status) is in fact independent of whether or not that relative was alive at the time of the child’s birth!

Yet again, we have another finding consistent with a genetically dominated mechanism of intergenerational status transferーif we expect factors such as educational expenditures, cultural transmission, introduction into patronage networks, etc. to play large roles, then whether or not a given relative was alive at all during someone’s childhood should factor into the degree of their influence on that person’s terminal status in later life! In contrast, if the causality is largely mediated through genetic relatedness, then this is exactly the result we would expect to obtain.

Going further, Clark mentions in the podcast that random deaths during the Spanish Flu of 1918 can be used as an exogenous shock for the presence or absence of a parent, and that the estimated effect of parental absence on long-term outcomes is remarkably small:

There’s almost no effect on people’s social outcomes from losing a parent. And it turns out there’s a nice study that was done in Sweden, looking at the 1918 influenza epidemic. And the losses of parents that came from that as a kind of a quasi-random intervention, which essentially finds the same thing for Swedish society that people losing a parent actually had surprisingly little impact on their lives and future.

It is a little unclear which study Clark is referring to specifically, but I did find Dribe et al. (2022), The Effect of Parental Loss on Social Mobility in Early Twentieth-Century Sweden:

We employ sibling fixed-effects models and the Spanish flu as an exogenous mortality shock to assess the importance of endogeneity bias in associations between parental loss and socioeconomic outcomes. Maternal death led to worse socioeconomic outcomes in adulthood in terms of occupational and class attainment, as well as for social mobility. The effects seem to be causal but the magnitudes were small. For paternal death, we find no consistent pattern, and in most models there was no effect on sons’ socioeconomic outcomes. The patterns were similar for sons and daughters and do not support the theory that parental loss had important negative effects on socioeconomic outcomes in adulthood.

This is a truly astonishing outcome. The idea that the wholesale loss of a parent in early childhood has minimal causal effect on future socioeconomic outcomes runs almost entirely contrary to any model that posits any contribution of parental or family environment on long-run outcomesーexactly opposite to what the vast majority of people would predict.

Finallyーwhat about public education? We all know that education has a beneficial effect on long-run outcomes, right? Well… in yet another paper, Clark and Cummins (2020), Does Education Matter? Tests from Extensions of Compulsory Schooling in England and Wales 1919-22, 1947 and 1972 show that cohorts which exogeneously received half an extra year of schooling in certain years (due to extensions of compulsory schooling) had zero effect on long-term outcomes such as health or income. Going even further, they examine the prior literature on these educational ‘impacts’ in some detail and claimーconvincingly in Milky Eggs’ opinionーthat the bulk of positive effects reported are due to a combination of publication bias and unsophisticated p-hacking!

The diversity of findings reported here, encompassing work both within and outside Clark’s own group, as well as the remarkable strength of overall consistency with a genetic hypothesis of talent and status transmission, is ultimately very suggestive of a truly minimal (at best) contribution of environmental factors to long-term outcomes.

Broader literature on intergenerational mobility

I am no expert on the economic literature on intergenerational mobility. Nevertheless, I believe it is valuable to attempt to put Clark’s findings into a broader context, at least with respect to the few other papers in the field with which I am familiar.

A notable, and recent, set of findings on American socioeconomic mobility come from Raj Chetty, who has made a career out of analyzing large, proprietary datasets such as IRS tax returns or the entire Facebook friend graph. For example, in Social capital I: measurement and associations with economic mobility, Chetty finds that:

The share of high-SES friends among individuals with low SES—which we term economic connectedness—is among the strongest predictors of upward income mobility identified to date^10,11. Other social capital measures are not strongly associated with economic mobility. If children with low-SES parents were to grow up in counties with economic connectedness comparable to that of the average child with high-SES parents, their incomes in adulthood would increase by 20% on average. Differences in economic connectedness can explain well-known relationships between upward income mobility and racial segregation, poverty rates, and inequality^12,13,14.

A naive reading of these results might have one believe they are contradictory with Clark’s broader narrative of a genetic undercurrent being the primary driver of intergenerational (im)mobilityーbut is that really the case here? Notice that “share of high-SES” friends is claimed to be one of the “strongest predictors of upward income mobility.” Previously, we noticed that matching of romantic partners assorted extremely strongly (r > 0.65) on underlying characteristics; it would only be natural to assume that matching for friends also assorts strongly, even if not as strongly, on intrinsic talent and intelligence. As such, one could argue that Chetty’s observations are in fact fully consistent with Clark’s work.

Chetty breaks down this enigmatic variable, “share of high-SES friends,” further in Social capital II: determinants of economic connectedness:

We show that about half of the social disconnection across socioeconomic lines—measured as the difference in the share of high-socioeconomic status (SES) friends between people with low and high SES—is explained by differences in exposure to people with high SES in groups such as schools and religious organizations. The other half is explained by friending bias—the tendency for people with low SES to befriend people with high SES at lower rates even conditional on exposure. Friending bias is shaped by the structure of the groups in which people interact. For example, friending bias is higher in larger and more diverse groups and lower in religious organizations than in schools and workplaces.

Here, though, it remains unclear to me that “underlying talent” is adequately accounted for in Chetty’s analysis. For example, could one not imagine that differences in “exposure to people with high SES” is, at least partially, an implicit measure of a given child’s parents’ intrinsic talent? Similarly, might “friending bias,” especially among “diverse groups” (potentially subject to quite strong selection effects on parental talent given the challenges of immigration into America) not be especially confounded with the genetic characteristics that determine later-life status?

To be sure, it is hardly the case that Chetty’s papers prove Clark’s hypotheses. I simply wish to point out that, despite their superficially egalitarian veneer, they are in fact consistent with Clark’s genetic interpretation of intergenerational mobility, and that, given the weight of other evidence on Clark’s side, the overall strength of the genetic interpretation is quite high.

Another interesting report on quite a different socioeconomic context comes from Alesina et al. (2020), Persistence Despite Revolutions:

Can efforts to eradicate inequality in wealth and education eliminate intergenerational persistence of socioeconomic status? The Chinese Communist Revolution and Cultural Revolution aimed to do exactly that. Using newly digitized archival records and contemporary census and household survey data, we show that the revolutions were effective in homogenizing the population economically in the short run. However, the pattern of inequality that characterized the pre-revolution generation re-emerges today. Almost half a century after the revolutions, individuals whose grandparents belonged to the pre-revolution elite earn 16 percent more income and have completed more than 11 percent additional years of schooling than those from non-elite households. We find evidence that human capital (such as knowledge, skills, and values) has been transmitted within the families, and the social capital embodied in kinship networks has survived the revolutions. These channels allow the pre-revolution elite to rebound after the revolutions, and their socioeconomic status persists despite one of the most aggressive attempts to eliminate differences in the population.

The major finding, in pictoral form:

Here, the authors claim that the “pre-revolution elite” were able to “rebound” after the Cultural Revolution, transmitting human and social capital across an entire generation to their grandchildren. Although this study does not claim to disambiguate the various mechanisms of transmission and types of transmitted capital, it is notable that they do not mention genetics as a contributory factor. Yet, in some sense, that omission makes the role of genetics starkly apparentーis it really plausible that after the wholesale destruction and redistribution of bourgeois wealth as well as decades of chaos and social upheaval, elite grandparents really managed to transmit nearly all of their “knowledge, skills, and values” as well as “social capital embodied in kinship networks” to their grandchildren in such a durable and timeless form that we return to nearly pre-Revolution levels of socioeconomic inequality? If anything, this is quite strong corroborating evidence for the nearly deterministic role of genetic inheritance in long-run outcomes, provided that one grows up in a modern capitalist economy rather than Communist China!

Closing thoughts

In the end, can we really know, with absolute certainty, how much genetics contribute to long-run success? Noーor perhaps I should say, not yet. However, Clark’s work is compelling evidence in favor of a genetic model of talent transmission for multiple reasons:

• It displays remarkable consistency across multiple centuries’ worth of heterogeneous data
• It successfully predicted a high degree of mating assortment on genetic intelligence, later validated by modern genomic studies
• It is consistent with work from other groups, even if those groups do not present their work as such

The “genetic hypothesis,” so to speak, is therefore a parsimonious and elegant theory which appears to be the best explanation for a number of disparate phenomena observed across different cultures and multiple centuries’ worth of data.