How much of "Ideas Getting Harder To Find" is due to declining researcher skill?

In a now famous 2020 paper Nicholas Bloom, Charles Jones, John Van Reenen, and Michael Webb document a puzzling trend: since 1940, research inputs have increased by 23-fold but productivity growth, which economists usually suppose is the output of research, has stayed constant or declined. They propose a simple explanation for this trend: research gets harder the more you know. Ideas get harder to find.

I've written about this paper before and offered alternative explanations for the steep decline in research productivity, but a new paper by Lorenz Ekerdt of the Census Bureau and Kai-Jie Wu at Penn State has a different theory with some interesting empirical tests to back it up. “Self-Selection and the Diminishing Returns of Research” claims that around half of the fall in research productivity documented by Bloom et al. is not due to a permanent difficulty in finding ideas, but rather due to tons of new and less productive researchers being hired into the industry.

The Model

Ekerdt and Wu notice that as part of the expansion of research inputs, the researcher share of the labor force has increased, meaning the population of researchers has grown much faster than the overall population. If the distribution of research skills within a population is constant, i.e nobel prize quality researchers are always only .01% of the population, then increasing the share of the population dedicated to research must change the average skill of researchers.

This is easy to understand in the extreme. Imagine only 1% of the population has the capacity to do any research at all and they’re all already working as scientists. If you then triple the research labor force to 3% of the population, you’d see rising research spending but constant research output, since all of the new workers are expensive but useless. Thus, research productivity falls precipitously.

In general, if the slice of the population who self-select into the research sector are the ones with the highest research ability, then expanding that slice necessarily means hiring less productive researchers. As you add in these less and less productive new workers, the average productivity of research goes down.

The Bloom (2020) method attributes all of this fall to a permanent difficulty in discovering new ideas, and predicts that research productivity will continue to fall at this rate far into the future. But in Ekerdt and Wu’s model, research productivity will stop falling as soon as the research labor force stops expanding, and long-run economic growth will be much higher than Bloom et al. would predict.

Should We Believe It?

But why should we believe that all the extra researchers added over the past several decades have lower productivity than the ones who were there before, and how much of the divergence between research spending and TFP growth can this explain?

The central piece of evidence is comparing the wages of people who switch industries in to or out of research. The idea here is that “marginal” researchers i.e the ones most recently added as the research labor force expands, are more likely to leave the industry. Thus, if you compare the wages of researchers who will leave the industry to ones who won’t, you can get an estimate of the difference in their wages and therefore their productivity. If the marginal researcher has lower wages and is less productive than the average one, then expanding the labor force must be lowering its average productivity.

What they find is exactly that. Researchers who leave the industry in 2015 have lower wages in 2013 than researchers who will stay, controlling for age, age-squared, gender, race, and year-fixed effects.

I’m not entirely sure how convincing this evidence should be. It sounds reasonable enough that the marginal worker is more likely to switch industries but not so obvious that one can comfortably rest their entire identification strategy on it. It’s also not clear how good wages are as a proxy for productivity in the research sector which is both influenced by price insensitive governments and universities and faces large externality problems where research productivity may not be rewarded.

Still, the idea that the recent rapid expansion in the research labor force has lowered average researcher skill is intuitive to me. It fits with evidence I covered a couple weeks ago that the recent rapid expansion in the college educated labor force has lowered average college graduate literacy significantly.

The So What

Carrying the empirical estimates through their model, Ekerdt and Wu claim that Bloom et al. greatly overestimated the difficulty of finding new ideas and thus underestimated long run economic growth rates by about half a percentage point. That compounds to big differences over long enough time periods.

This sounds optimistic but we can't wash our hands and think that research is actually going fine. This result is still a sign of a big problem, just a different one. The problem is not in inexorable feature of research that just makes it harder the more we know, instead it’s a talent allocation problem.

In the author’s model, the highest productivity researchers are automatically sorted into the research sector. But in real life there are enough nobel prize winners who are certain they wouldn’t get tenure today and brilliant PhDs who are sent back to their home countries to be confident that the expansion of research to low productivity marginal workers is due to poorly-designed and prestige-obsessed scientific funding mechanisms rather than a lack of available talent.

Research productivity is falling because of higher education’s failure to raise skills and because of science funding’s failure to select high productivity researchers. The massive expansion of education and research funding since WWII has failed to produce any increase in productivity growth rates. This is not mostly because science just got a lot harder.