Does An Aging Population Actually Lead to Slower Economic Growth?
In a previous post we looked at the economic arguments for immigration based on population dynamics, which suggested that a younger population is likely to lead to a larger and more productive labor force and therefore greater economic output, leading to a faster growing economy.
This argument feels intuitive and squares with what we would generally expect, but a question is whether we can actually measure the impact of an aging population on economic growth.
Accurately measuring the effects of aging on growth is important, especially given the demographic trends of the current US population where we expect to see an aging population over the next few decades.
If an aging population leads to slower growth, we will need policy interventions to counteract that; however, if it does not then the need to act to address the aging population specifically is less of a concern.
A priori there are reasons for why an aging population might actually lead to faster economic growth.
Acemoglu and Restrepo, two economists, argue in a paper that an aging population has not correlated with slower economic growth using cross country evidence, finding it is positively correlated with growth.
They posit that the reason is due to robots: firms have increasingly relied on machines to perform tasks that normally younger workers would be doing. This substitution has in turn led to productivity increases that have compensated for the aging of the population by making existing workers more efficient.
Conversely, Nicole Maestas, Kathleen J. Mullen, and David Powell, three economists, argue in a working paper that an aging population does lead to slower economic growth, leveraging evidence across US states through a smaller and less productive labor force.
So who is right? Does aging lead to slower growth through a smaller and less efficient labor force? Or does it lead to capital substitution of labor and therefore faster growth?
The Acemoglu and Restrepo paper correlates the change in the ratio of young to old workers (defined as the ratio in the population above 50 to the ratio between 20 and 49, which is known as age dependency) from 1990 to 2015 with the change in GDP per capita over that same time period across countries.
As the graph above shows, there is a slight positive relationship as opposed to the negative relationship one might expect.
The authors of the paper also examine this relationship using regressions that allow them to control for various factors like initial GDP in 1990, differences across countries (through fixed effects), and test alternative measures of aging (like average age of the population above 20).
All these alternative specifications find either a significant and positive relationship or an insignificant relationship (that is sometimes positive or negative depending on the model).
One cautionary note is that these regressions are not causal estimates. To compensate for this the authors attempt to use IV regressions that could elucidate the causal relationship between aging and economic growth.
Columns 4 and 6 in the table above show the results of their IV regressions, which are either positive or not significant, matching the observational results.
The instruments they use are prior birthrates, which in theory should correlate with the age structure of the population as higher birthrates in previous years would affect the future age structure.
However, their instruments are rather weak (as indicated by the first stage results), and they use birthrates from only as early as 1965 as an instrument. That would leave people born then around 50 at the end of when they predict economic growth rates, which doesn’t really capture aging of the workforce sufficiently because those people are probably still employed and not retiring en masse.
Still, the authors take this overall as a sign that aging has not led to slow economic growth and propose another mechanism that could explain the results: scarcity of workers has led firms to adopt new technology to compensate for a small and/or less productive workforce.
They show that the change in robots per million hours worked correlates with the change in ratio of old to young workers across countries, meaning that countries with older workforces substituted more capital for labor.
The Maestas et. al paper comes to the opposite conclusion, finding that population aging leads to a large, negative impact on economic growth. They don’t leverage cross country data though, instead using data from across US states.
Their specification of aging is also slightly different. Instead of focusing on the ratio of old to young, they focus on the fraction of the state’s population above 60 years old. They also use time series panel data with multiple observations per state (as opposed to single observations for each country over a period like Acemoglu and Restrepo), predicting the growth in GDP ten years down the road for each decade from 1980 to 2010.
Interestingly, the basic correlation across US states is negative instead of positive, showing differences from the basic cross country results.
To account for the fact that the data is observational and therefore won’t directly reveal the causal impact of aging on growth, the authors also use instrumental variables.
They use the age distribution in a given year to predict the subsequent decades’s age distribution using national trends, effectively forecasting what we would expect the age distribution to be in the future. Specifically, they use the fraction of people above age 50 to predict the next decade’s fraction of people above 60.
Intuitively, if a state today has lots of people above 50, then in 10 years it should have lots of people above age 60 as the 50 year olds age to 60 year olds.
Using this they perform regressions similar to Acemoglu and Restrepo, looking at the effect of change in GDP over time against the change in fraction of the population above 60 with their instrument. They also test different specifications.
Consistently across time periods and specifications the authors here find significant and negative effects of aging on population growth.
So how do we square these two results?
It is hard to compare them because they use different methodologies and different data sources (cross country vs. cross state). To facilitate comparison we can apply the Maestas et al. methodology to the cross country data that Acemoglu and Restrepo use. Note we use the instruments from Maestas et al. to avoid the issues of Acemoglu and Restrepo.
Here are the observational results using the specification of Maestas et al. on the cross country data of Acemoglu and Restrepo.
We see that the point estimates are insignificant unless we include country fixed effects, which suggests that the Acemoglu and Restrepo results potentially came about because they did not properly account for cross country trends over time. Once we include those, the negative and significant effect of aging on growth is robust to definition of aging, inclusion of year trends, and weighting by population.
The instrumental variable results where we use the expected old fraction of the population as an instrument for actual old fraction of the population confirm the observational findings as well.
We have a strong first stage regression and a significant and negative coefficient in the IV regression, suggesting a negative causal relationship between aging and growth.
These results overall would seem to suggest that the Agemoglu and Restrepo specification did not properly account for differential cross country trends nor specify an appropriate IV estimation strategy to accurately estimate the effect of aging on growth and that the estimates from the Maestas et. al paper are more realistic.
This would confirm the idea that an aging population does indeed lead to slower economic growth, but there is one caveat. Acemoglu and Restrepo argue that capital substitution of robots for workers will slow or even reverse the impact of an aging population and their work showed a positive correlation between aging workers and robot adoption.
One could imagine that if firms were sufficiently forward looking, they would anticipate a demographic slowdown in the labor force and invest in acquiring robots before it happens. This would make the predicted age distribution in the population not a valid instrumental variable because it would correlate to GDP through both actual population demographics and robot adoption.
Hence, we still need more work to truly identify the full causal impact of aging on growth. The weight of the evidence here though suggests there is very likely a negative effect but one whose magnitude could be diminished through capital substitution of labor with robots by forward looking firms.
Data Sources:
Data obtained from the World Bank website.
