There's obviously much more in the paper, including multiple robustness exercises and a cool solution to classical measurement error. You can read it at the link and the replication files are available on GitHub.
papers.ssrn.com/sol3/papers....

This punts the question to why is there a high elasticity of substitution between graduates and non-graduates? I argue it's because many graduates and non-graduates have overlapping skillsets. Many graduates work in occupations that don't require a degree.

One reason is that graduates and non-graduates were always more substitutable than the canonical race between education and technology model implied. Those results came from non-stationary time series regressions and differencing out the unit root makes that result disappear.

I get null results that are precise enough to rule out the results one might expect from previous research at the aggregate level. Therefore, the elasticity of substitution must be relatively high. Why is that?

I answer this question by grouping majors into skill groups using hierarchical clustering. Then, I construct a pseudo-panel to run major fixed effects and IV regressions of responsive wages are to the supply of similar majors.

Previous research had suggested that US graduates and non-graduates are not very substitutable (Race Between Education and Technology). We also know that the returns to a degree vary massively by subject. How substitutable are graduates from different majors?

And gay men have more stable marriages than straight couples!

There are some meta analyses that suggest the effects here across studies are quite low after accounting for publication bias though. onlinelibrary.wiley.com/doi/full/10....

I think what's happened here is that OLS results are likely too large on labour market and health outcomes and the RD estimates standard errors suggests the RD is heavily underpowered for detecting the OLS results.

Sadly, not very. onlinelibrary.wiley.com/doi/full/10....