Interesting, thanks for the review (pun intended). I was thinking of adding a deep literature search option -- currently it does a simple one through @perplexityai.bsky.social, but at some additional cost that can be upgraded to a deep lit search. Do you think that'd be valuable?
No idea what that means lol
The project's spirit is very close to that of @chicagohai.bsky.social @chenhaotan.bsky.social: AI-assisted reviewing should be open and not for profit. Go check out their AI reviewer @ openaireview.org/index.html as well! If it's better, use that one, or submit proposals to create the best of both!
Again, the whole system is open-source (github.com/Davidvandijc...) so there are no hidden costs, and you can check what the system does by looking at the source code.
My hope is that many people will contribute code and submit issues to make this open system as good as possible!
As the underlying engine, you can choose among any AI model that's on @openrouter.bsky.social (which is nearly all models!).
Less budget? Use any open-source model for <$1!
For the best results, use SOTA models like Claude Opus 4.6, GPT-5.4, etc., at slightly higher cost.
Setup is easy: coarse.ink/setup
Create an account at @openrouter.bsky.social , buy a couple $$ worth of credit, and obtain your API key. You can also log in directly with your OpenRouter account.
Reviews remain active for 90 days
None of your data is stored beyond that, and all calls to whichever AI model you use are done with an explicit denial of data collection, so Claude/Gemini/ChatGPT/... cannot retain your data or use it for training
See: coarse.ink/privacy
The review is provided in an interactive panel with major and minor comments that you can trace in the text and tick off when done, see an example here
coarse.ink/review/79c43...
Some details on the performance comparisons to other popular systems are here coarse.ink/compare
These are all judged by Gemini-3.1-pro, so caveats apply.
Anecdotally, some of my colleagues and friendly ppl on X said it rivals reviews from Refine.Ink, when using a SOTA model.
I coded up an open-source, not-for-profit AI paper reviewer that rivals the performance of
@reviewer3com.bsky.social, Refine.ink, and Stanford Agentic Reviewer (according to Gemini 3.1). Costs <$2!
Live @ coarse.ink. Plug in paper, @openrouter.bsky.social key, and email. #econsky
Another "good" week on the US academic job market #econsky #econbluesky
davidvandijcke.com/joe_tracker/
Thanks for sharing!
Thanks for sharing!
US academic economics market continuing to pull away from the COVID market this week.
I created a little tracker for JOE here for those who want to play around with the data. Updates weekly: davidvandijcke.com/joe_tracker/ #econbluesky #econsky
The paper integrates causal inference, functional data analysis, and optimal transport, developing (FAST!) new tools for empirical researchers.
If you use micro data or focus on inequality effects, I’d love to discuss potential applications! #EconTwitter
(11/11)
The method opens up a large new class of RDDs for distributional policy evaluation.
E.g.: local minimum wage impacts on wage distributions, district ed reforms on grade distributions, close elections on constituent outcomes...
For more details:
(10/) arxiv.org/abs/2504.03992
The results suggest a classic "equality-efficiency tradeoff" under Democratic governors:
Incomes at the top 10% of the distribution drop significantly, but this effect weakens and becomes statistically imprecise lower down the distribution.
(9/)
Finally, I illustrate the method's use in a close-election RD (or rather, R3!) design.
I study how Democratic vs Republican governors affect families' income distributions within their states when they barely won/lost their election.
(8/)
I validate both methods through extensive simulations, which show rapid convergence to the quantile treatment effects...
...unlike existing quantile RD methods, which do not converge (but remain useful in the classic setting!)
(7/)
I develop uniform confidence bands and data-driven bandwidth selection for both approaches, which are fully implemented in an R package (available at davidvandijcke.com/R3D).
(6/)
To estimate this unknown beast, I propose two closely related estimators.
One extending local polynomial regression to random quantiles, and a functional version of that, based on local Fréchet regression (which has better mathematical and computational properties).
(5/)
I propose a new concept of treatment effects for R3D: Local Average Quantile Treatment Effects.
Instead of averaging over conditional scalar outcomes, they average over conditional distributions!
This captures the average distributional shift across the cutoff.
(4/)
Standard RD methods can't handle such settings, as they don't account for the two levels of randomness—within and across distributions.
R3D solves this problem by modeling outcomes as random distributions rather than random variables!
(3/)
The method is useful when aggregate units receive treatment, but your outcome varies within the unit.
E.g., firms receive a subsidy when their revenue (X) drops below a cutoff, and you want to study this subsidy's effect on the employee wage distribution
(2/)
Hi Bluesky!
I'm excited to share my job market paper (for the 2025-26 market)!
It introduces a new extension of RDD where outcomes are entire distributions: Regression Discontinuity Design with Distributions (R3D).
Thread below 👇 (1/)