Innovators in big cities have long benefited from stronger “local knowledge spillovers” as compared to innovators in smaller cities and towns. But various evidence suggests this advantage has been eroding for decades.
(In jargon-free language, the notion is that innovators in cities are physically close to more people, and therefore have better access to more knowledge/ideas in those people’s heads)
The canonical paper establishing the importance of local knowledge spillovers is Jaffe, Trajtenberg, and Henderson (1993). JTH ask us to consider three patents: A, A*, and B. A and A* are technologically similar patents, and B is a patent cited by A but not A*. JTH show A and B are more likely to belong to the same city than A* and B. In other words, patents are more likely to cite local patents, relative to non-citing controls.
This study has been cited over 9,000 times, including more than 5,000 times in the last decade. But it predates the internet era(!). What’s happened since 1993?!
Kwon, Lee, Lee, and Oh (2019) give JTH a much needed update. They replicate JTH’s methodology but bring data up through 2015 (and with much larger samples of patents). They find local knowledge spillovers are strengthening!
Or are they?
Kuhn, Younge, and Marco (2019) have an alarming paper (for us innovation researchers) that shows patent citations don’t mean what they used to. Their paper is really interesting and warrants more discussion at a future date - for now, suffice it to say, there has been a sharp increase since 2000 in the number of citations of highly dubious quality. To demonstrate the importance of these changes, KYM do a quick and dirty update of JTH through 2010, using both the standard set of patent citations and a selected sample more likely to reflect genuine knowledge flows (the “improved selection”).
At left, you can see the probability a citing and cited patent belong to the same city increased from 3.9% to 6.4% between 1975 and 2010 using a conventional set of citations, but only increased from 4.3% to 4.7% when you use the improved sample of citations. At right, you can see relative to control patents the increase in the probability a citing and cited patent belong to the same city has actually declined significantly, especially when you used the improved selection. In 1975, a citing and cited patent were 3x as likely to belong to the same city as a control and the cited patent. By 2010, this was under 2x. In other words, patents are a lot less likely to cite other local patents than they used to be.
That said, there’s not a lot of detail on KYM (the paper is focused on changing citation quality, not local knowledge spillovers) so let’s be cautious with the citation evidence.
Mewes (2019) takes an alternative approach, and assumes innovation is about discovering new ways to combine existing knowledge. Access to knowledge about different technologies locked away in other people’s heads is quite useful for this kind of innovation. And indeed, a paper by Berkes and Gaetani (2019) finds city centers do generate more unusual combinations of technologies.
Mewes’ work uses patents, but not patent citations so it shouldn’t have the same data quality issues. Instead he uses the technology classifications assigned to patents; when a patent is classified as belonging to technology classes X and Y, Mewes interprets this as an indicator that the patent combined pre-existing technologies X and Y. He is particularly interested in novel and unusual combinations of technology, since we would expect cities to have disproportionate advantage for these novel combinations. Mewes looks at how the size advantage of big cities in producing “atypical” combinations has changed between 1850 and 2010.
The figure above calculates the scaling exponent for county population over 1850-2010. The interpretation is, that if the scaling exponent is x, then a 2-fold increase in population leads to a 2^x-fold increase in atypical/typical technology combinations. The higher is x, the bigger the advantage of a big population in generating unusual combinations. But the scaling exponent peaked in the 1970s and has been on a steady decline since. Big cities no longer have as strong an advantage in generating unusual combinations of technology.
Finally, Packalen and Bhattacharya (2015) take a different tack yet again. They scan the text of all patents and pull out important one-, two-, and three-word sequences (e.g., “microprocessor” and “polymerase chain reaction”). They call these word(s) “concepts” and interpret them as technological ideas. Because they can observe the date each concept is first mentioned in a patent, they can measure the “age” of the idea.
They then look at whether patents in cities use newer ideas than patents from outside cities. They find the patents of big cities used to have a much higher probability of mentioning a very young concept, but this faded over time. By the 2000s, they can no longer rule out zero effect.
So, by a couple different measures (still all based on patents) the advantage innovators get for locating in big cities has eroded, at least via one channel. In one sense, this is hardly surprising, since the internet drastically lowers the cost of accessing distant knowledge. We should hardly expect it to be otherwise.
But in an another sense, this is quite surprising given the increasing concentration of the tech sector into a smaller and smaller number of superstar cities. Why would more tech workers move into high-cost cities, even as one of the benefits of doing so decreases? Well, for the first and second studies, there still are stronger local knowledge spillovers in big cities - just not as strong as in the past. Given the winner-take-most dynamics of tech, it may make sense to pile into cities to grab that (smaller) advantage. At the same time, if costs in those cities continue to mount, and the strength of local knowledge spillovers continues to decline, it may be that the trend towards hyperconcentration reverses in the future.
Thanks for reading! If you like this, you can help to improve this newsletter by sending me interesting papers on the economics of innovation, especially stuff you think isn’t well known.