Is your future income written in your DNA?

A newly published paper makes the claim that variation in people’s income can be partly traced to variations in their genes. Indeed, it identifies over a hundred specific genetic variants that are statistically associated with income in a large sample of people derived from the UK Biobank. To some, this idea is frankly preposterous – a naïve and outrageous over-reach of genetic determinism and reductionism, with strains of social Darwinism. To others, it is completely expected – not trivial, in terms of the work involved, but certainly not at all surprising and not so earth-shattering in terms of social implications.

The devil is in the details, of course, of the methodology and the results, and, importantly, the way they are presented and interpreted.

The idea that something like a person’s income could be partly heritable – that is, that variation in income across the population could be partly attributable to genetic differences between people – is in fact, not new at all and really not remarkable at this stage. Decades of research in behavioral genetics has clearly shown that pretty much every human behavioral trait and every life outcome – from income to marriage to divorce to educational attainment to going to prison – is partly heritable. This is because these experiences are partly driven by our psychological traits, which are themselves partly driven by genetic differences.

Given the existence of these genetic effects, it may be interesting to try and figure out the underlying mechanisms. That starts with identifying some of the actual genetic variants that are associated with the metric of interest in some population, and then figuring out the causal mechanisms that link them to the phenotype. I say “may be” interesting because it’s not actually obvious that it always is – sometimes the details of the genes involved are very informative and others times they’re really not. Of course, you can’t know until you look.

The paper in question, by W. David Hill and colleagues, describes the results of a genome-wide association study (GWAS) of income in a large sample of over 280,000 people from the UK (all ethnically “white”). GWAS analyse sites in the genome where there is a common variant – the DNA “letter” at that position may be an “A” in some people and a “C” in others, for example. The test is simply to see if the frequency of those variants differs in people with high income versus low income. (In the same way that the frequency of smoking varies in people with lung cancer versus without). GWAS look at millions of such sites (known as “SNPs”) across the genome across the whole sample of hundreds of thousands of people. As a result of the huge sample, these studies can detect even small differences in frequency and differentiate significant ones from random noise.

It was already well established prior to this study that income is partly heritable. What the authors wanted to do was identify some of the causal genetic variants, see what other traits or life outcomes might be mediating that causal connection, and illuminate the types of genes and biological processes involved. They set out their causal model in Figure 1 of their paper:

This figure shows that the idea that income is partly genetically driven is not so preposterous after all. It follows directly from the fact that intelligence is strongly heritable, that intelligence is a contributing factor in educational attainment, and that intelligence and educational attainment together contribute to what kind of job someone gets and what they earn. None of these relationships is deterministic or complete – they are all just partial correlations, but they’re real and robust. When the causal pathway is shown like that, it would be amazing if income were not partly heritable!

(Note: That doesn’t mean that that diagram is necessarily correct or complete, however. It makes a lot of assumptions, one of which we know is violated. More on that below).

The abstract sets out the findings of the paper:

Socioeconomic position (SEP) is a multi-dimensional construct reflecting (and influencing) multiple socio-cultural, physical, and environmental factors. In a sample of 286,301 participants from UK Biobank, we identify 30 (29 previously unreported) independent-loci associated with income. Using a method to meta-analyze data from genetically-correlated traits, we identify an additional 120 income-associated loci. These loci show clear evidence of functionality, with transcriptional differences identified across multiple cortical tissues, and links to GABAergic and serotonergic neurotransmission. By combining our genome wide association study on income with data from eQTL studies and chromatin interactions, 24 genes are prioritized for follow up, 18 of which were previously associated with intelligence. We identify intelligence as one of the likely causal, partly-heritable phenotypes that might bridge the gap between molecular genetic inheritance and phenotypic consequence in terms of income differences. These results indicate that, in modern era Great Britain, genetic effects contribute towards some of the observed socioeconomic inequalities.

So, what does all this mean? Let’s take a look at the results themselves, the conclusions based on them, and the broader implications.

The results

The main general findings are that income is partly heritable and that variation in a measure of intelligence (results on a cognitive test) partly mediates those genetic effects. I think we knew both of those things already, and they’re certainly not surprising, but the authors may rightly claim that this is a more explicit demonstration than has been made previously. Because they identify some associated variants they are able to use somewhat different approaches (such as genetic correlations and Mendelian randomisation) to try and tease out the causal relationships. 

I’m a little skeptical of those kinds of analyses in this context, or at least think the conclusions we can draw from them are somewhat limited. The idea of genetic correlations is that when you find some genetic variants associated with one phenotype (such as income) you can then see whether those genetic variants are also associated with another phenotype (such as intelligence) and measure how strongly the effects of the variants correlate across the two phenotypes. In this case, the authors find a strong genetic correlation (rg=0.69) between intelligence and income, and comparable genetic correlations between income and health and longevity.

That’s fine, but that number can be misunderstood. It doesn’t imply that the actual overall correlation between intelligence and income is that strong. It doesn’t imply anything about that correlation in fact. The genetic variants collectively only “explain” a fraction of the variance in each of the two phenotypes. That effect may well be highly correlated across the two phenotypes, but still very minor and not sufficient to drive a strong overall correlation between the two phenotypes.

You see a lot of these genetic correlations in the literature these days, because they’re easy to do, especially in databases with rich and diverse phenotyping, like the UK Biobank. The results can definitely be interesting and informative, but it’s worthwhile keeping the primary effect sizes in mind when interpreting them.

The authors also use a statistical technique called Mendelian randomisation to try and extract a causal relationship from what is otherwise just correlative data on genetic effects on intelligence and income. The application of this technique here rests on some assumptions, the main one being that variants that affect intelligence (which could thereby affect income) do not also affect some other traits that could have independent effects on income. The authors argue against this possibility of  “horizontal pleiotropy” in their discussion of limitations of the study, but it seems very difficult to rule it out. That’s not to say there’s really any reason to doubt that model, as it is inherently plausible, and a wealth of data shows that intelligence partially correlates with future earnings.

Beyond those general points, the main novelty of the paper lies in the identification of specific SNPs that are associated with the phenotype of income. As mentioned, many of these are also associated with intelligence, which is argued to be the causal mediator. When you find an associated SNP (or multiple ones in a small region) you can often deduce what gene is likely affected by those genetic variants. The authors identify 144 genes in this way that harbour genetic variants statistically associated with income. 

What kinds of genes are involved?

So, what kinds of genes are they? What are their normal functions? Well, the first thing to note is that they are highly enriched for genes expressed in the nervous system. That’s good news, methodologically speaking, because it provides a crucial reality check. One danger in performing GWAS for a social phenotype like income is that that average income might vary geographically across the country. (In fact, we know it does for the UK). This means you could pick up SNP associations that are really tracking local differences in ancestry, rather than anything in a real causal chain to the phenotype.

The authors use a number of statistical methods to control for that problem but there’s always a worry that such measures are insufficient. However, if the GWAS signals were just due to cryptic population stratification you would not expect enrichment for any specific classes of genes – just signals all over the genome with no biological meaning and possibly no biological effect at all. So, the enrichment for neural genes strongly argues that the signals detected are real and also fits with the favoured explanation that intelligence is a causal mediator.

Beyond that, however, there’s not really a lot that can be said about these genes that is particularly illuminating. First, the associated genes were not enriched for any particular biochemical functions. Second, the authors use a variety of gene expression datasets to look for any enrichment of associations in genes with higher expression in particular parts of the brain or cell types. And they do come up with some statistically significant enrichments, but, to be frank, as a neuroscientist, I don’t know what to make of that information. I think probably nothing.

For example, the associated genes show enrichment for “expression differences in cerebellar hemisphere, at Brodmann area 9 (BA9) of the frontal cortex, the nucleus accumbens and at Brodmann area 24 of the anterior cingulate cortex (BA24)”. These findings rely on databases of gene expression profiles across the brain. But most genes are expressed in many different areas (many in all). Some are a bit higher in some areas than others, but that does not mean their function is restricted to the areas with higher expression. Those statistical descriptors do not reflect the true richness and dynamics of gene expression in the brain. And the data presented in this paper are statistical enrichments of those statistical enrichments (i.e., there’s a bit more signal in genes that are expressed a bit more in those areas). It is highly doubtful that the genetic effects on income (via intelligence or other psychological traits) are mediated in any specific way (or even predominantly) by changes in those highlighted brain areas. 

The same can be said for the data on cell type enrichments:

“Cell-type-specific analysis revealed that the expression that was specific to the serotonergic neurons and to medium spiny neurons was associated with income. Medium spiny neurons have previously been linked to schizophrenia, which has a strong cognitive component and has previously been linked to glutamatergic systems, including the N-methyl-D-aspartate receptor signalling complex. Medium spiny neurons are a subtype of GABAergic inhibitory neurons. Future work should examine if, like other cognitive traits, income is linked to both GABAergic and glutamatergic systems.”

This is a strange way of working round to the idea that maybe the two major classes of neurons in the brain are involved somehow. In any case, the same problems apply as with the brain area expression data – the real patterns of gene expression are simply not that specific. Cell types are defined by combinatorial profiles of gene expression, and most individual genes are expressed all over the place in many different cell types, in complex and dynamic ways.

I don’t mean to be overly critical of the attempts. These kinds of analyses might have thrown up some specific information – they certainly have for other kinds of phenotypes. But for complex psychological phenotypes, which really reflect the function or performance of the whole brain, we should not expect any anatomical or cell-type specificity, and my interpretation of the data presented in this paper is that, while there are some statistically significant signals, they are not biologically meaningful.

Finally, the authors integrate various sources of data to focus on a subset of the genes that seem most likely to have genetic variants directly affecting expression in the brain. They say: “These 24 genes therefore should be prioritized in follow-up studies”. It’s not obvious what such follow-up studies might entail. The effects of each of the variants on the phenotypes of interest are individually almost negligible, and provide no real experimental purchase. Of course one can look at the effects of the genetic variants on expression of the genes, and one can perform various studies to try and figure out what the normal functions of those genes is, at a biochemical or cellular or developmental level.

But that is a very different question from asking what are the effects of genetic variation in the gene. These effects can be very non-specific and not really related to what you would say the function of the gene is. The genomic program of brain development involves lots of genes with specific developmental functions but also relies on maybe ten-fold more genes with much more generic functions – not really part of the instructions of brain development per se, just general stuff required for everything to go right (like metabolic enzymes, for example). Most of the genetic variants affecting brain development will thus be doing so indirectly and non-specifically (and non-informatively).

So, I am not convinced that the identification of these associated genes actually sheds much light on the underlying biology driving the association with income (via intelligence or otherwise), except to reinforce the idea that those kinds of psychological traits are really emergent properties of the whole neural system and can’t necessarily be tied to specific molecular systems.

The conclusions

The final line of the abstract states: “These results indicate that, in modern era Great Britain, genetic effects contribute towards some of the observed socioeconomic inequalities.” This is clearly the intended take-home message. And it is true, in a literal sense, (if the main results are methodologically sound, as they seem to be), but also a bit misleading and open to misinterpretation. A more accurate statement would be that “genetic effects contribute a minor amount towards some of the observed socioeconomic inequalities”.

Based on their calculations, the total heritability of income that can be tagged by common genetic variants in the population sample under study is only 7.4%. And using polygenic scores based on the results of the GWAS to predict income in an independent sample only explained 2% of the variance. Given that, the unqualified wording of the abstract leaves open the possibility of people interpreting the effects as larger than they really are. (And then running with that interpretation to make all kinds of other claims).

Note that the overall genetic effect may in fact be a good bit larger, as twin studies have estimated the total heritability of income at 40-50%. These studies account for non-genetic familial effects, in that these should be shared equally between twins regardless of whether they are monozygotic or dizygotic. The observation is that pairs of MZ twins are much more similar in income than DZ twins, implying a genetic effect. A likely explanation for why the current study could not account for all of that heritability is that much rarer genetic variants are responsible for much of it.

However, it is extremely important to emphasise that the heritability is not a fixed number. There is no right answer. It only measures the proportion of phenotypic variance that can be attributed to genetic differences within the population sample actually being studied. It cannot be generalised to other populations or taken as a sweeping statement of truth. Variation in a social measure like income can obviously be hugely dependent on social and cultural factors – much more so in some settings than others. If you take a very culturally homogeneous sample, where the variation in such social and cultural factors is low, then the heritability of the trait will be correspondingly higher (because it is a proportion). But across more diverse social samples, the heritability could be much lower. And heritability within one population says nothing – at all – about the origins of differences between populations.

A concern is that people will take only the widest, most sweeping interpretation of the results described here to bolster the argument that socio-economic success is simply meritocratic, in the sense that it is determined in some simple way by people’s innate intellectual abilities and aptitudes. (See here for an example of such an argument). The reality is vastly more complex. Not only are environmental factors hugely important, but they interact with genetic factors in complex, non-linear ways.

This is where the authors’ basic model as presented in Figure 1 is too simple, in my view. The models used to partition variance assume that the effects of genetic and environmental factors will be independent of each other and will simply add up. It’s difficult to determine whether or not the assumption of additivity actually holds in a study like this, but it seems likely that the effects of some environmental factors will vary depending on a person’s genetic make-up, and vice versa. But more importantly, there are correlations between a person’s genes and the environment they experience. These arise because they share genes with their parents, who partly create the environment for their children. I think a more accurate causal pathway diagram would look like this:

The authors do say that they cannot rule out such dynastic effects (gene-environment correlations across generations) as a possible confounder and refer to another paper that suggests they may be especially important in analyses of the genetics of educational attainment. The problem is the magnitude of such effects is extremely difficult to calculate in the population study design employed.

This means the straightforward causal pathway – where the genetics of the individual drives the heritability of income – is likely too simple. The real scenario is sure to involve complex positive feedback loops that entangle and amplify the effects of genetic endowments and social capital over generations.

Finally, the authors add these important caveats on how their results should be (and should not be) interpreted:

“A further limitation is that molecular genetic analyses of phenotypes, such as intelligence, income or SEP, appear prone to being misinterpreted. Such misunderstandings include describing associated variants as genes for income, or the misinterpretation that any associated variant, and indeed any nonzero heritability estimate, is evidence for genetic determinism or the immutable nature of these phenotypes via environmental intervention. We include a figure (Fig. 1) that illustrates that genetic variants do not act directly on income; instead, genetic variants are associated with partly heritable traits (such as intelligence, conscientiousness, health, etc.), which have their own complex gene-to-phenotype paths (including neural variables) and are ultimately associated with income. Therefore, the genetic variant–income associations discovered here are no more for income than they are for these other traits. For more discussion of the implications of these results, aimed at the general reader, we have provided a Frequently Asked Questions (FAQ) document in Supplementary Note 2.”

The FAQs are very useful, but it seems unlikely that many readers will actually even notice they exist or go to the trouble of digging them out, meaning that despite these efforts by the authors, these findings are sure to be misinterpreted and misrepresented. I do think some of the plainer language and important disclaimers in the FAQ answers could have been included in the main text to better pre-empt misunderstandings.

I’m sure this paper will continue to attract a lot of attention and we will see more studies like it (like this one, for example) that explore the potentially fraught areas of genetic effects on complex social outcomes.


Popular posts from this blog

Undetermined - a response to Robert Sapolsky. Part 1 - a tale of two neuroscientists

Grandma’s trauma – a critical appraisal of the evidence for transgenerational epigenetic inheritance in humans

Undetermined - a response to Robert Sapolsky. Part 2 - assessing the scientific evidence