Correlation is not causality. It is perhaps one of the most frequently used phrases when I write about research here, and I have noticed that it can sometimes even irritate people. But a new NBER working paper by Andrew Johnston beautifully illustrates that the story usually only begins there. Johnston’s question is simple: can working from home cause people to have more children? That idea is not so crazy. Working from home means less or no commuting, can provide more flexibility, and might make it easier to balance work and family. But is that the case?

There was already previous research pointing in that direction. In an international survey of adults aged 20 to 45, the estimated number of children over their lifetime was 0.32 higher when both partners worked from home at least one day a week than when neither did. But that is precisely where our problem lies. Perhaps people are having more children because they work from home. Perhaps people who do want children look more often for work they can do from home. Correlation, causality, you know the drill by now.

Johnston tries to circumvent that problem cleverly. He looks not only at who is working from home, but also at what people studied. For every field of study, he calculates how suitable the professions graduates entered before 2020 were for working from home. Computer science, for example, scores much higher than nursing. Quite clever, because the advantage is clear: most people had already chosen their field of study long before the pandemic began. They could hardly have chosen a specific program in, say, 2010, because it would have offered them more opportunities to work from home during the as-yet-unknown pandemic of 2020.

And then there was that gigantic, natural experiment in 2020. People from degree programs well-suited to working from home suddenly started working from home much more often. For women, one standard deviation more teleworkability after 2020 meant a 3.9 percentage-point higher likelihood of working from home.

But… did they have more children afterwards? No, they had just slightly fewer! One standard deviation more teleworkability was associated with a decline in the annual birth rate of approximately 1.9 per cent. Notably, the effect was primarily among married women. Johnston found no significant difference among unmarried women.

At the same time, something else happened. Married women in more teleworkable fields of study were more likely to work full-time. The author estimates that this increase in full-time work could account for approximately half of the decline in births. Perhaps, therefore, working from home makes it easier to combine work and family, but does this simultaneously mean that women are more likely to continue working full-time, resulting in a higher opportunity cost for children?

But does that prove that working from home leads to people having fewer children? No, because what was that fantastic natural experiment again that suddenly caused working from home to increase? That’s right: a global pandemic.

And of course, that pandemic did much more than just send people to their home offices or simply confine them at home. It also affected professions and sectors very differently. Just think of nurses. Their education scores very low on opportunities for working from home, while their working conditions also changed considerably during the pandemic, to put it mildly.

Johnston is very well aware of this problem himself. He explicitly calls it a pandemic shock that affected fields of study differently for reasons other than working from home, citing it as a threat to his causal interpretation. He therefore conducted several additional analyses. If he omits education programs, little changes. It becomes more interesting when both health and education programs are omitted. The association remains negative, but among married women it becomes about half as large and is no longer statistically significant. However, one should not make too much of this either: this new estimate does not differ significantly from the original. Welcome to the wondrous world of causal inference.

There is another problem, however. Fewer births between 2021 and 2024 do not necessarily mean fewer children overall. Perhaps people are simply postponing having children. Johnston finds a 3.1 per cent decline among married women aged 35 to 39. For them, catching up later will naturally be more difficult. But the author also acknowledges that we will have to wait longer before we can say anything about the final number of children in the US.

And right at the end stands perhaps the most important limitation: this expansion of working from home occurred during a pandemic. Working from home, when introduced under normal circumstances, can have a different effect, according to Johnston. At the same time, he rightly notes that the decline in births continued to grow through 2023, when most pandemic restrictions had already been lifted.

I therefore consider this primarily a nice lesson on correlation and causality. Not because the researcher confuses the two. On the contrary. He makes a considerable effort to get closer to causality, but in doing so demonstrates how difficult it is.