The Molecular Genetic Architecture of Self-Employment

Think about what it would take to hire for entrepreneurship. Risk tolerance, independence, drive, creativity, and the willingness to bet on yourself. Now imagine trying to find those traits in someone's DNA — scanning fifty thousand people's genomes, variant by variant, looking for the biological signatures of that choice. Van der Loos and colleagues did exactly that, and they came up almost completely empty. What that near-emptiness tells us is the real story here. Entrepreneurship is not an obvious candidate for genetic research. It feels like a decision, not a disease. But the motivation for studying it genetically is real and grounded. Income, education, and occupation all connect to health — there is a consistent inverse relationship between socioeconomic status and cardiovascular disease, and occupational choice has been linked to coronary heart disease risk in women. Many of these outcomes are themselves partly heritable, which raises a pointed question: are the same genetic factors shaping both the economic choices we make and the health outcomes that follow? Entrepreneurship sits at this crossroads. Previous research links it to higher stress and poorer average health outcomes but also to greater job satisfaction and life satisfaction. Understanding its genetic architecture could eventually inform both economics and public health. For the study, van der Loos and colleagues operationalized entrepreneurship as self-employment — having started, owned, and managed a business — because it is the measure most consistently available across large genotyped populations. They used two complementary strategies. The first was a classical twin analysis using the Swedish Twin Registry, computing tetrachoric correlations that model a continuous underlying tendency rather than a binary yes or no within identical and fraternal twin pairs. They then fitted ACE models that partition variance into additive genetic effects, shared environment, and individual environment. The second strategy was molecular: using the method developed by Yang and colleagues, implemented in software called GCTA, which estimates genetic relatedness between unrelated individuals from genome-wide single nucleotide polymorphism data — SNP for short, meaning a single-letter difference in the DNA sequence — and then relates that estimated relatedness to phenotypic similarity. On top of that, they ran a genome-wide association study, or GWAS, meta-analysis across sixteen separate cohorts. The scale of the effort is worth pausing on. The GWAS meta-analysis pooled 50,627 participants of European ancestry — 7,734 who had ever been self-employed and 42,893 who had not. Each cohort independently imputed their genotype data to a common reference panel and ran association tests across roughly 2.4 million SNPs. The results were then combined using a fixed-effect weighted meta-analysis. This was, at the time of publication, the first large-scale molecular genetic study of any economic variable framed this way. Here is what they found. The twin analysis estimated that about 55 percent of the variance in the tendency toward self-employment is attributable to additive genetic effects — 67 percent for males and 40 percent for females. The molecular SNP-based estimate came in at 25 percent of phenotypic variance explained by common autosomal SNPs jointly, a statistically significant result in the pooled sample. So roughly half of the twin-based heritability appears to be captured by common genetic variants working together. Then the GWAS ran and found nothing genome-wide significant. Not a single SNP crossed the threshold of a p-value less than five times ten to the negative eighth. The strongest pooled signal was a variant near a gene called RNF144B, with a p-value of about four times ten to the negative sixth — still more than an order of magnitude away from significance. The top male-specific signal, in a gene called HECW2, reached one and a half times ten to the negative seventh. Fifty-eight suggestive SNPs with p-values below ten to the negative fifth were carried forward into a replication sample of 3,271 individuals. None replicated. A previously reported candidate variant in the dopamine receptor gene DRD3 also failed to hold up. Gene-based tests across roughly 17,700 genes — including genes specifically proposed in the literature as entrepreneurship candidates — returned nothing significant either. The prediction results are the starkest summary. Polygenic scores built from the discovery meta-analysis and tested in an independent sample explained a maximum of 0.18 percent of variance in self-employment. Less than one-fifth of one percent. The p-value on even that tiny result just barely cleared 0.039 — it squeaked through significance. Sex-specific scores showed no association at all. The gap between a twin heritability of 55 percent and a predictive score explaining under 0.2 percent is not a contradiction — but it is striking, and it demands an explanation. That explanation is what the authors call a highly polygenic architecture. Think of it this way: rather than a handful of genetic variants each having a meaningful effect on the probability of becoming an entrepreneur, there may be hundreds or thousands of variants, each nudging the odds by a tiny fraction of a percent. In aggregate, they add up to the heritability the twin studies detect. But in the discovery phase of a GWAS, with finite sample size, each individual effect is estimated with too much noise to be pinned down reliably — and a polygenic score built from noisy estimates predicts poorly even when the underlying signal is real. This pattern is not unique to self-employment. It mirrors what researchers have found for other complex traits. For height, the most studied polygenic trait, twin studies estimate heritability at around 80 percent, but Yang and colleagues found that common SNPs explain roughly 45 percent of variance in height. Intelligence, personality, and several common diseases show similar gaps. Self-employment's numbers — 55 percent twin heritability, 25 percent SNP-based heritability, and nothing significant at the individual variant level — place it squarely in that same family of outcomes. The null result is not a failure of the trait to have a genetic basis. It is a signal about the structure of that basis. The study's power calculations make this concrete. The meta-analysis had roughly 80 percent power to detect a common variant — minor allele frequency of 0.25 — with an odds ratio of about 1.11 in the pooled sample. That is a modest effect size. Nothing with even that moderate an effect turned up, which places a real empirical ceiling on single-variant effects in self-employment: if any common SNP had an odds ratio of 1.11 or larger, this study almost certainly would have found it. Several limitations the authors acknowledge are worth holding onto. Self-employment is measured with single-item binary questions across diverse cohorts — someone who becomes self-employed one year after the assessment counts as a control, and someone who reluctantly freelances to survive is grouped with the ambitious founder. That phenotypic noise reduces power. Pooling sixteen studies across different countries and time periods adds heterogeneity. And the trait itself is biologically distal — self-employment is the endpoint of a long causal chain that runs through personality, risk preferences, social environment, opportunity, and chance. Van der Loos and colleagues suggest that future work should target more biologically proximate endophenotypes: risk preferences directly measured, confidence assessed psychologically, and independence as a stable personality dimension. These are closer to whatever biological mechanisms might exist, and GWAS studies of those traits might find variants that the self-employment measure cannot detect. What the genome appears to say about the choice to start a business is this: the propensity is real, it is partly written in DNA, and it is written very small, in very many places. The heritability is genuine — identical twins are meaningfully more similar in their entrepreneurial behavior than fraternal twins, and that pattern holds across multiple samples. Common SNPs collectively carry about half of that signal. But no single variant matters enough to be named, and no score built from current data can meaningfully predict who will take the leap. The biology is real. It is just diffuse — statistical rather than deterministic, a broad current rather than a switch. That may be the most important finding of all. The same genome that makes you 0.001 percent more likely to start a company also makes ten thousand other small contributions to who you are. The trait is heritable the way almost everything interesting about human behavior is heritable — pervasively, fractionally, and far beyond the resolution of any single study to decode. This lecture was created by ennepō. Go to https://ennepo.ai to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.