The Most Misread Statistic in Psychology

Heritability is the proportion of variation in a trait, within a particular population at a particular time, that is attributable to genetic variation.

Every clause in that sentence is doing work, and the popular uses of the word drop most of them. It is about variation, not about the trait. It applies to a population, not to a person. And it is indexed to a time and a set of conditions, so it changes when those change.

Three misreadings follow, and they are everywhere. That a heritability of 0.6 means sixty per cent of a person's trait is caused by genes. There is no coherent statement of that form, because the number is a ratio of variances across a group. That a heritable trait cannot be changed, needing glasses is highly heritable and glasses work perfectly; phenylketonuria is entirely genetic and is managed completely by diet. And that heritability measures how genetic something fundamentally is, rather than how much of the variation in one place at one moment tracks genes.

The consequence that most reliably surprises people is what happens when environments are equalised. Make schooling perfectly uniform and the remaining differences in educational outcome would be more genetic, so heritability would rise towards one: in a society that had just made the largest environmental intervention imaginable. High heritability can therefore be a marker of equal conditions rather than of fixed destiny, which is close to the opposite of how the statistic is usually deployed in argument.

The cleanest illustration is agricultural and worth carrying around. Take genetically varied seed and grow it in perfectly uniform soil: all the variation in height is genetic, and heritability approaches one. Take genetically identical seed and grow it in varied soil: heritability is zero. The plants are unchanged in both experiments. What differs is the distribution of environments, and that is all the statistic has ever described.

There is a further complication that makes the neat partition of variance considerably less neat. Gene-environment correlation: people partly select and shape the environments that suit their propensities. A child inclined towards reading is given more books, seeks out more reading, and is placed among people who expect it. The resulting difference is scored as genetic by the arithmetic and is delivered through environments that those propensities helped to produce. The statistics cannot easily separate the two, and this is one of the field's genuinely hard problems rather than a quibble.

None of which makes heritability a useless measure. It is a real quantity, carefully estimated, and it tells you something about the population it was computed in. What it will not do is any of the work it is usually recruited for: predicting an individual, establishing what is fixed, or settling an argument about whether an intervention is worth attempting.