The recent publication of a study of the genetics of educational attainment is once again raising questions and controversies about the potential use of biological information to inform education policy and practice. In the BioEduDataSci project, funded by the Leverhulme Trust, we have been collecting and examining a very large sample of texts and commentary to understand the development of the new genetics of education and its potential consequences. Such studies are characterized by being highly data-scientific, focused on identifying minute molecular differences, and highly controversial both ethically and scientifically. This research has a high profile in the media, and draws criticism from a variety of scientific and political perspectives.
The new study, conducted and published by the Social Science Genetic Association Consortium (SSGAC), is no different. It is the fourth in a series of studies linking DNA differences to educational outcomes. The first educational attainment study (EA1) examined a sample of around 100,000 people, the second (EA2) a sample of 300,000, and the third (EA3) featured a sample in excess of 1 million. EA4, however, features a sample of more than 3 million genotyped individuals, and has identified more than 3,000 tiny genetic variants—known as single nucleotide polymorphisms (SNPs)—that are said to be associated with years spent in education. The paper, published in Nature Genetics, is extremely technical, but its key findings have been usefully summarized by one of its main investigators, and a follow-up çommentary paper by one of the authors stating simply that ‘genes matter when it comes to educational performance and social outcomes’.
Like previous projects and publications linking DNA to educational outcomes, the EA4 study certainly raises serious ethical challenges. These include the lack of representation in the data analysed, the potential to reinforce existing racial categories and discriminatory outcomes, produce negative self-fulfilling prophecies, and possible appropriation by right-wing conservatives and ideologically-motivated scientists to make racialized, hard-hereditarian arguments about IQ and social stratification. The horrific history of eugenics in education, particularly through the intelligence testing movement, has left a dark legacy that means current studies of the genetics of education come under particularly intense scrutiny, not least because of the persistence of discriminatory practices and policies grounded in genetic theories of difference. The term ‘genetics of education’, widely used in popular accounts of recent research, may itself problematically harden assumptions about the association of genes with educational achievement or even IQ.
Rather than rehearse those issues here, however, I want to focus on one specific scientific dispute in relation to genetic educational attainment studies. The post draws from our wider research detailing the organizational, conceptual, methodological and technical systems and structures underpinning the new genetics of education, and exploring the knowledge claims and proposals for intervention emerging from such studies. What we are interested in this post is not so much developing an external critique of educational genetics, but tracking some of its internal conflicts and their implications.
New genetics of education
Over the last 15 years, new studies of the genetics of education have appeared from two overlapping fields of scientific inquiry. Behaviour genetics is a branch of psychology examining the genetic bases of human behaviours and traits, or how genotypes influence phenotypes (like cognitive ability). It has recently adopted high-tech and data-driven genomics methods to study the ‘molecular genetic architecture’ of ‘educationally-relevant traits’ and phenotypes, including cognitive ability and intelligence.
Meanwhile, social science genomics, or sociogenomics, represents the interdisciplinary combination of genomics with certain branches of sociology, economics, and political science. It is interested in the biological structures and mechanisms that interact with environmental factors to produce socioeconomic outcomes. Educational attainment is taken by sociogenomics as a key socioeconomic outcome that is related to other outcomes such as occupation, social status, wealth, and health.
Where behaviour genetics and sociogenomics overlap is in their methodologies and techniques of analysis. They both use highly data-intensive research infrastructure, such as biodata repositories and data mining software, and methods that can identify vastly complex associations between thousands of minute SNP variations and their correlation with educational outcomes or relevant traits.
They use methods such as genome-wide association studies (GWAS), processing enormous quantities of SNP data, and calculating ‘polygenic scores’—a quantitative sum of all SNP variant data—and apply them to education. Thus, educational attainment studies such as those undertaken by the SSGAC use GWAS methods and SNP analysis to produce polygenic scores for education, which can then be used to predict the attainment of independent samples. Similar approaches have been taken to a range of education-relevant phenomena, such as intelligence and learning.
The SSGAC’s huge-sample educational attainment studies are taken as gold standard models for investigating the genetics of education by both behaviour genetics and sociogenomics researchers. Their findings underpin the arguments in Robert Plomin’s controversial book Blueprint, where he proposes that DNA data could used as the basis for personalized ‘precision education’, are explored in detail by Katherine Paige Harden in The Genetic Lottery to underpin her argument for educational reforms, and animate a wide range of other sociogenomics studies (although the SSGAC reports no practice or policy implications from the EA4 study due to being only ‘weakly predictive’).
Not all scientists agree, however, that multimillion-sample genomics studies offer any useful insights into the genetics underpinning education, even less that they might inform how education itself is organized.
Perhaps surprisingly, one of the most consistent and vocal critics of data-driven genomics studies of education is a leading behaviour geneticist, Eric Turkheimer. Turkheimer certainly believes in the heritability of human behaviours—that is, that a certain portion of behaviour is influenced by genetics rather than being entirely environmentally shaped. He wrote the ‘three laws’ of behaviour genetics affirming as much. A fourth law was added in 2015 (by authors from the SSGAC) stating that ‘a typical human behavioural trait is associated with very many genetic variants, each of which accounts for a very small percentage of the behavioural variability’.
Polygenicity among hundreds of thousands or even millions of SNPs has become a defining law of social and behavioural genomics studies, including those focused on the genetic heritability of educational behaviours and outcomes. Leading scientists claim these methods promise to ‘open the black box of heritability’ and finally reveal the pathways from DNA to social outcomes such as educational achievement.
Turkheimer, however, has become highly critical of the methodological turn to genomics when studying complex behaviours and outcomes, including in the SSGAC EA studies, as his useful recent tweet-response to EA4 indicates. The basis of the critique is the so-called ‘missing heritability problem’. Turkheimer argues in a new paper with Lucas Matthews that the concept of ‘heritability’—an ‘estimate of the proportion of phenotypic [behavioural] variance that is statistically associated with genetic differences’—has changed with shifting methodologies for its measurement.
Of course, heritability estimates are historically constructed and contested. Basically, earlier forms of behaviour genetics utilizing quantitative genetics methods, such as twin studies, identified DNA to play a large influence on any behavioural trait, often in the region of 50-70%. For example, in relation to education, quantitative behaviour genetics found around 50% of variation in intelligence, as measured by IQ tests, was heritable.
In contrast, Matthews and Turkheimer argue that recent high-tech, data-intensive genomics methods, such as genome-wide association studies (GWAS), have hugely increased computational power but reduced explanatory power. For example, ‘cutting-edge GWAS have recently estimated that only 10% of variance in IQ is statistically associated with differences in DNA’. They point out that studies of educational attainment from the SSGAC are therefore scientifically underwhelming, as they account for somewhere in the region of 12-15% of variance, despite the huge samples and computational power put to the analysis. This gap is what’s known as the missing heritability problem.
The issue for Matthews and Turkheimer is that most solutions to missing heritability appear to be focused on throwing even bigger samples and more processing power at the problem. They describe this as ‘dissolving the numerical gap’ between traditional quantitative and molecular computational kinds of heritability estimates, but argue ‘resolving the numerical discrepancies between alternative kinds of heritability will do little to advance scientific explanation and understanding of behavior genetics’. They note that ‘most writing on the topic expresses optimism that this day will soon come as researchers collect larger datasets and develop more sophisticated statistical genetic models of heritability’.
By contrast, Matthews and Turkheimer ‘argue that framing the missing heritability problem in this way—as a relatively straightforward quantitative challenge of reconciling conflicting kinds of heritability—underappreciates the severe explanatory and methodological problems impeding scientific examination and understanding of heritability’.
More urgent than closing the numerical gap, for them, is the persistent ‘prediction gap’, or the challenge of making accurate and reliable prediction from DNA to behaviour, and, even more so, the ‘mechanism gap’, which refers to the gap in explaining the specific mechanisms linking molecular genotypes to behavioral phenotypes. There remains, they suggest, a ‘black box’ of hidden mechanisms that simply solving the numerical gap will never discover.
These gaps in prediction and explanation of mechanisms are especially acute for studies of the genetics of education. They pose a challenge to claims that genetic data could be used—in the not so distant future—to open the ‘black box of heritability’, or even inform educational policy or practice in schools:
the putatively causal relationship between … SNPs and differences in educational outcomes is entirely opaque, other than the very general assertion that many of the SNPs are close to genes that are expressed in neural tissue. Until scientists have identified, described, and substantiated causal-mechanical etiologies that would explain why countless SNPs are correlated with behavioral outcomes like IQ and educational attainment, then what we call the mechanism gap of the missing heritability problem remains a daunting and persistent scientific challenge. … Highly complex human behavioral traits and outcomes such as intelligence and educational attainment are farthest from dissolution: the numerical gaps, predictions gaps, and mechanism gaps for these cases may never be resolved.
The paper highlights two important issues confronting the new genetics of education. First, despite significant investment in scientific infrastructure, data analytics technologies, and high-profile publications, educational genetics studies remain a long way from opening the ‘black box’ of the specific genetic mechanisms that underpin educational outcomes. And second, it highlights how educational genetics studies are not just ethically controversial and subject to external critique, but scientifically controversial and internally contested too.
Making biodatafied realities
Another key part of Matthews and Turkheimer’s argument is that the heritability estimates produced by quantitative genetics are of a very different kind than the heritability models produced by molecular genomics methods such as GWAS and SNP analysis. This is not just a matter of quantitative innovation, then, but a qualitatively different mode of investigation which produces a very different kind of knowledge.
In this sense, Matthews and Turkheimer seem close to suggesting that the computational infrastructure of genomics makes a significant difference to the knowledge that is produced. This is an argument familiar in critical data and infrastructure studies, where it is assumed data infrastructures are far from merely neutral interfaces to access factual reality, but instead represent ‘expressions of knowledge/power, shaping what questions can be asked, how they are asked, how they are answered, how the answers are deployed, and who can ask them’.
Likewise, the epidemiologist Cecile Janssens has recently argued that polygenic scores, in particular, emerged as a ‘pragmatic solution’ to the statistical problem of calculating very large SNP associations in genomics. Technically, a polygenic score is calculated using particular data-mining software applications, computing formats, algorithms and statistical standards created by specialists in high-tech genomics research laboratories. As Janssens suggests, this level of technical mediation in the construction of polygenic scores matters.
Polygenic scores, Janssens argues, ‘do not “exist” in the same way’ as other measurable biological processes such as blood pressure, but only as ‘algorithms’, ‘models’, or ‘simplifications of reality’. Her concern is that polygenic scores, as pragmatic solutions to a statistical problem, might create a new ‘biological reality’ and be used as the basis for certain forms of intervention, despite being only simplified models.
Janssens’ concern, like that of Matthews and Turkheimer, reflects important critiques of genomics in science and technology studies. The historian of bioinformatics Hallam Stevens, for example, argues that ‘these algorithms and data structures do not merely store, manage, and circulate biological data; rather, they play an active role in shaping it into knowledge’.
Polygenic scores and heritability estimates produced through informatic forms of biology, then, may generate very different conceptions of what constitutes ‘biological reality’. As Joan Fujimura and Ramya Rajagopalan have argued elsewhere, data-intensive genomics ‘is ultimately a statistical exercise that depends on the analytic software itself and the information that goes into the statistical software’. As such, they emphasize, when a GWAS or SNP analysis reveals complex associations, ‘these underlying patterns are known only through the data and data producing technologies of the geneticists’.
If Matthews and Turkheimer are correct, then it seems like the kind of heritability of educational attainment that recent GWAS, SNP and polygenic score studies have uncovered is a different kind of heritability that can be known only through the data and technologies of educational genetics research.
As a pragmatic solution to a computing problem, polygenic scores are changing the ways that DNA is understood to affect social and behavioural outcomes. They are based on a kind of heritability that exists primarily as a computational artefact–one which, despite its weak predictivity and absence of explanatory power, is nonetheless attracting significant media and public attention as a way to understand the molecular bases of educational outcomes.
Even while overall effect sizes of such studies may remain modest, they may already be shifting public and professional discourse towards a more biological perspective, infusing educational debates with the vocabulary of genotypes, heritability, phenotypes, and genetic architectures, despite the persistent explanatory gap in the underlying science.
The social life of bio-edu-data science
What we can suggest here, then, is that new biological knowledge and even biological realities are being created through the use of data-intensive genomics technologies and methods in educational genetics research. As Matthews and Turkheimer argue, a different kind of heritability emerges from the complex scientific and data infrastructures of GWAS and polygenic scoring. For Janssens, polygenic scores only exist as algorithms, not as embodied substance. This challenges assertions that once the polygenic ‘genetic architecture’ underpinning educational outcomes is objectively known through big biodata analysis, it may be possible to design educational interventions on the basis of such knowledge.
Rather, new knowledge about the genetics of education is generated through distinctive computational systems, software, and methods that all leave their mark on understandings of the genetic substrates of educational outcomes. What emerges from such studies, as Matthews, Turkheimer and Janssens suggest, are new bio-edu realities that have been produced through computers, data processing, and particular statistical applications, rather than simply unmediated insights into the objective molecular substrates that underpin students’ educational achievements.
But these new realities can be consequential despite being scientifically contested or only weakly explanatory. They may animate enthusiasm for ideas about genetically-informed schooling, and could potentially lead to a hardening of biological explanations, sometimes dangerously motivated by racism, for the complex social, political, cultural and economic factors that shape students’ achievement in schools.
For these reasons, in the BioEduDataSci project, we’re tracking the ‘social life’ of the new genetics of education. This means trying to understand the social contexts and conditions of new knowledge production, the reception of such new knowledge, and its social, political and ethical implications as such knowledge circulates in the media and in public, often by being translated and interpreted and sometimes turned to harmful results.
The new genetics of education is not yet a settled science, and if critical voices–even within fields like behaviour genetics, as well as its critics–are correct, may never be. Nonetheless, the new genetics of education remains a fast-moving science in the making, surfacing complex issues and problems that need addressing and debating among many stakeholders across the biological and social sciences, policy and practitioner sites, before any emerging findings are considered as insights for implementation.
UPDATE: A few days after posting this, the EA3 educational attainment study was cited to justify an act of horrific racist violence resulting in many deaths in Buffalo, USA. This has animated urgent calls among the genomics community for scientists of such research to take far more responsibility for their study findings – or perhaps not even publish it at all given its dangerous consequences – generating counterarguments about scientific freedom.