Paul Goldsmith-Pinkham
@paulgp.com
Yale SOM professor & Bulls fan. I study consumer finance, and econometrics is a big part of my research identity. He/him/his
Wrote a new paper on the econometrics of financial event studies, would value feedback! It's very new.
With my amazing grad student Tianshu Lyu www.tianshulyu.com, who is on the market. You should hire him!
paulgp.com/papers/finan...
With my amazing grad student Tianshu Lyu www.tianshulyu.com, who is on the market. You should hire him!
paulgp.com/papers/finan...
November 19, 2025 at 4:43 AM
Wrote a new paper on the econometrics of financial event studies, would value feedback! It's very new.
With my amazing grad student Tianshu Lyu www.tianshulyu.com, who is on the market. You should hire him!
paulgp.com/papers/finan...
With my amazing grad student Tianshu Lyu www.tianshulyu.com, who is on the market. You should hire him!
paulgp.com/papers/finan...
They're using PSID data
November 18, 2025 at 7:27 PM
They're using PSID data
This paper claims housing is not homothetic: trevorcwilliams.github.io/files/submis...
November 18, 2025 at 7:26 PM
This paper claims housing is not homothetic: trevorcwilliams.github.io/files/submis...
We also tackle inference. Our argument: if assignment is truly random at the individual level, heteroskedasticity-robust (non-clustered) SEs are appropriate. The assignment mechanism should guide your inference. And UJIVE's default SEs properly account for many-instrument corrections.
November 17, 2025 at 3:06 PM
We also tackle inference. Our argument: if assignment is truly random at the individual level, heteroskedasticity-robust (non-clustered) SEs are appropriate. The assignment mechanism should guide your inference. And UJIVE's default SEs properly account for many-instrument corrections.
We show a nice test for average monotonicity using UJIVE, similar to the original Kitagawa test, but valid under many instruments and controls. doi.org/10.3982/ECTA...
November 17, 2025 at 3:06 PM
We show a nice test for average monotonicity using UJIVE, similar to the original Kitagawa test, but valid under many instruments and controls. doi.org/10.3982/ECTA...
The standard monotonicity assumption (all examiners rank cases identically) is probably too strong, as Imbens and Angrist (1994) pointed out in their original paper. But "average monotonicity" (Frandsen et al. (2023)) is more plausible and sufficient.
November 17, 2025 at 3:06 PM
The standard monotonicity assumption (all examiners rank cases identically) is probably too strong, as Imbens and Angrist (1994) pointed out in their original paper. But "average monotonicity" (Frandsen et al. (2023)) is more plausible and sufficient.
The paper also guides how to think about heterogeneous treatment effects in many IV settings. It turns out that you can write the overall estimand as a set of pairwise IV regressions where the weights are proportional to the squared leniency distances between pairs of examiners.
November 17, 2025 at 3:06 PM
The paper also guides how to think about heterogeneous treatment effects in many IV settings. It turns out that you can write the overall estimand as a set of pairwise IV regressions where the weights are proportional to the squared leniency distances between pairs of examiners.
The UJIVE estimator gets the order right: residualize the instruments first, then do leave-one-out estimation. This properly constructs relative leniency without own-observation contamination. With 100+ fixed effects, this difference between JIVE and UJIVE matters.
November 17, 2025 at 3:06 PM
The UJIVE estimator gets the order right: residualize the instruments first, then do leave-one-out estimation. This properly constructs relative leniency without own-observation contamination. With 100+ fixed effects, this difference between JIVE and UJIVE matters.
(As an aside, we have a really lovely Appendix section that formally combines all these results together, mainly thanks to Michal, which I really encourage you to read if you like metrics!)
November 17, 2025 at 3:06 PM
(As an aside, we have a really lovely Appendix section that formally combines all these results together, mainly thanks to Michal, which I really encourage you to read if you like metrics!)
It turns out having many controls also creates this issue – intuitively, this makes sense that the same “own observation” problem shows up in estimation for control variables just like for excluded instruments. And in this setting, JIVE is actually biased in the opposite direction!
November 17, 2025 at 3:06 PM
It turns out having many controls also creates this issue – intuitively, this makes sense that the same “own observation” problem shows up in estimation for control variables just like for excluded instruments. And in this setting, JIVE is actually biased in the opposite direction!
This problem can be solved – the jackknife estimator was designed with this problem in mind, and excludes its own observation in estimation. But, in many leniency settings there are additional subtleties.
November 17, 2025 at 3:06 PM
This problem can be solved – the jackknife estimator was designed with this problem in mind, and excludes its own observation in estimation. But, in many leniency settings there are additional subtleties.
Holding fixed the sample and growing the number of examiners, your IV estimates will get pulled toward OLS. The bias can be substantial - this is Angrist & Krueger (1991), and Bound, Jaeger, and Baker (1995) all over again.
November 17, 2025 at 3:06 PM
Holding fixed the sample and growing the number of examiners, your IV estimates will get pulled toward OLS. The bias can be substantial - this is Angrist & Krueger (1991), and Bound, Jaeger, and Baker (1995) all over again.
This logic extends (with caveats) with many examiners and even many controls. But the subtleties of estimation become more important.
The easiest way to see this is by realizing this setting with many examiners is exactly a many-weak-instrument setting, a classic econometrics problem.
The easiest way to see this is by realizing this setting with many examiners is exactly a many-weak-instrument setting, a classic econometrics problem.
November 17, 2025 at 3:06 PM
This logic extends (with caveats) with many examiners and even many controls. But the subtleties of estimation become more important.
The easiest way to see this is by realizing this setting with many examiners is exactly a many-weak-instrument setting, a classic econometrics problem.
The easiest way to see this is by realizing this setting with many examiners is exactly a many-weak-instrument setting, a classic econometrics problem.
Fortunately, patent examiners are randomly assigned within units, and they decide whether to grant the patent. Turns out, according to Farre-Mensa et al. (2019), that there is a lot of variation across examiners in their approval rates!
November 17, 2025 at 3:06 PM
Fortunately, patent examiners are randomly assigned within units, and they decide whether to grant the patent. Turns out, according to Farre-Mensa et al. (2019), that there is a lot of variation across examiners in their approval rates!
Take a setting with patent applications at the USPTO. We’re interested in estimating the effect of patent approvals on subsequent innovation and economic success. We could run OLS of our outcome y (say, subsequent innovation) on x (patent approval)
November 17, 2025 at 3:06 PM
Take a setting with patent applications at the USPTO. We’re interested in estimating the effect of patent approvals on subsequent innovation and economic success. We could run OLS of our outcome y (say, subsequent innovation) on x (patent approval)
Leniency designs are very popular, but with real implementation issues that can meaningfully affect your results. Here's a list from Frandsen et al. (2023) with just a subsample of applications where it's used.
Here's what we found and what you should do differently.
Here's what we found and what you should do differently.
November 17, 2025 at 3:06 PM
Leniency designs are very popular, but with real implementation issues that can meaningfully affect your results. Here's a list from Frandsen et al. (2023) with just a subsample of applications where it's used.
Here's what we found and what you should do differently.
Here's what we found and what you should do differently.
New paper with @instrumenthull.bsky.social and Michal Kolesár on leniency/judge IV designs.
This article is targeted for the Journal of Economic Perspectives, and so we wrote it with the goal of being accessible to a wide range of users, including advanced undergrads!🧵
arxiv.org/abs/2511.03572
This article is targeted for the Journal of Economic Perspectives, and so we wrote it with the goal of being accessible to a wide range of users, including advanced undergrads!🧵
arxiv.org/abs/2511.03572
November 17, 2025 at 3:06 PM
New paper with @instrumenthull.bsky.social and Michal Kolesár on leniency/judge IV designs.
This article is targeted for the Journal of Economic Perspectives, and so we wrote it with the goal of being accessible to a wide range of users, including advanced undergrads!🧵
arxiv.org/abs/2511.03572
This article is targeted for the Journal of Economic Perspectives, and so we wrote it with the goal of being accessible to a wide range of users, including advanced undergrads!🧵
arxiv.org/abs/2511.03572
This was @instrumenthull.bsky.social 's joker moment:
November 15, 2025 at 8:41 PM
This was @instrumenthull.bsky.social 's joker moment:
Update your syllabus and stay on the frontier - it will increase your students’ wages. Epic work by my colleagues @barbarabiasi.com and @profsongma.bsky.social #linkoftheday
www.barbarabiasi.com/uploads/1/0/...
www.barbarabiasi.com/uploads/1/0/...
November 15, 2025 at 12:44 AM
Update your syllabus and stay on the frontier - it will increase your students’ wages. Epic work by my colleagues @barbarabiasi.com and @profsongma.bsky.social #linkoftheday
www.barbarabiasi.com/uploads/1/0/...
www.barbarabiasi.com/uploads/1/0/...
“Candidates whose research is more similar to the committee are more likely to win” - we are just horrible at hiring in academia IMO #linkoftheday
academic.oup.com/ej/advance-a...
academic.oup.com/ej/advance-a...
November 15, 2025 at 12:40 AM
“Candidates whose research is more similar to the committee are more likely to win” - we are just horrible at hiring in academia IMO #linkoftheday
academic.oup.com/ej/advance-a...
academic.oup.com/ej/advance-a...
A skydiving man captured in front of solar activity - actually real! #linkoftheday
www.reddit.com/r/spaceporn/...
www.reddit.com/r/spaceporn/...
November 14, 2025 at 2:42 PM
A skydiving man captured in front of solar activity - actually real! #linkoftheday
www.reddit.com/r/spaceporn/...
www.reddit.com/r/spaceporn/...
24hrs ~ 84k searches
November 14, 2025 at 2:42 AM
24hrs ~ 84k searches
Trying to make blogging a thing again
November 13, 2025 at 4:01 PM
Trying to make blogging a thing again
Pleased with this: created an automatic crawler for my website that aggregates posts from my Bluesky feed into daily "Links of the Day": paulgp.com/2025/11/12/l...
Basically trying do MR's daily posts but without all the annoying snark...
Basically trying do MR's daily posts but without all the annoying snark...
November 12, 2025 at 8:30 PM
Pleased with this: created an automatic crawler for my website that aggregates posts from my Bluesky feed into daily "Links of the Day": paulgp.com/2025/11/12/l...
Basically trying do MR's daily posts but without all the annoying snark...
Basically trying do MR's daily posts but without all the annoying snark...