David Van Dijcke
@packlesshepherd
Followers
851
Following
5K
Media
35
Statuses
146
PhD @umichECON. Academic Visitor @bankofengland. Alum @lmhoxford, @LeuvenEconomics. 🇧🇪🇺🇲. Views mine. Apple says my research draws 'inaccurate conclusions'.
Ann Arbor
Joined April 2016
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/)
11
91
613
I now exclusively want to write papers called "X can be faster/better/smarter".
man, scientists working on optimizing matrix multiplications have oppenheimer level of aura. - use a RL agent to spit out heckload of bilinear products.- slap two MILP to combine and filter those.- iterate on top of a Large Neighborhood Search flow until it’s fast fast. what the
3
0
10
RT @evajanster: New! Efficient Estimation of Nonlinear DSGE Models, w/ Sean McCrary.📄 We propose a broadly app….
0
22
0
RT @I_Am_NickBloom: *Call for papers for the 2025 Remote Work Conference*. Stanford, October 22 to 24 2025. In person presentations, and st….
0
30
0
FDR isn't just for #StatsTwitter #EconTwitter! Useful for #MachineLearning, #DataScientists, #climate, urban studies, #finance, gene expression, #marketing & more. If your data has jumps, FDR can help! 🚀. @EmoryEconomics @umichECON
0
1
15
Findings: A 25-35% drop in economic activity at the shutdown boundary! Much larger than prior estimates. This shows how vital reliable digital infrastructure is. Internet expansion benefits are gradual, but shutdown shocks are immediate & severe in digitized economies.
1
0
5
Using FDR, we:. Mapped actual shutdown-affected areas (not just admin lines). Estimated economic drop-off magnitude at these new boundaries using mobile data.
1
0
3
📄 APPLICATION: . We estimated the economic impact of an internet shutdown in Rajasthan, India (2021) 🇮🇳. This event created sharp, unknown geographic breaks in connectivity – a perfect test for FDR!
1
0
4
We also confirmed our theoretical results in simulations, showing fast convergence to:.1️⃣ The true jump locations.2️⃣ The true jump sizes.3️⃣ The true underlying regression surface!
1
0
3
That's the theory, but it works in practice too! . We implemented FDR in PyTorch so that it can solve super-fast (~60 seconds) on your computer's GPU.
1
1
4
FDR's Edge:.1️⃣ Unified: Jump locations & sizes are direct outputs. 2️⃣ Guaranteed: Global optima, formal ID & consistency for jumps/surfaces. 3️⃣ Realistic: Handles random design, correlated noise.
1
0
2
The original Mumford-Shah is non-convex. FDR uses "calibrations" for a convex version. This means global optima, reproducible results, & strong statistical guarantees! ✅
1
0
3
FDR builds on the famous Mumford-Shah functional (from computer vision) with a statistical twist. It handles random point clouds & correlated noise – vital for real-world data. 🌐
1
0
3
Many methods separate smoothing & boundary detection, or lack formal guarantees for jump location & size (esp. in multi-D). FDR estimates both, with guarantees, in one step & any dimension.
1
1
4
Why FDR? Multivariate jumps in data (policy impacts, market segments, tipping points) can offer key insights into the underlying problem structure. But finding them accurately in noisy, multi-D data is hard 🤯 . Example: internet disruptions during a shutdown in Rajasthan 👇
1
1
3
🔥New method: Free Discontinuity Regression (FDR)! (w/ Florian Gunsilius). Combines ✨multidimensional✨ changepoint estimation & nonparametric regression in one shot. Find jump locations & sizes in any dimension, w/ formal convergence guarantees🧵👇 . #EconTwitter #StatsTwitter
1
37
213
RT @JohnHolbein1: #RDDers check it out. Regression Discontinuity Design with Distribution-Valued Outcomes. "This article introduces Regre….
0
57
0
Great blog post on my JMP, IV, and interflex! Thank you for the nice words and fun read @BeatrizGietner.
We have a new post! 🙋♀️Which was a delight to write 🥰.With papers/guides by @instrumenthull, @packlesshepherd, @xuyiqing, and coauthors .🔗
1
1
5