Thibaut Lamadon

343 posts

Thibaut Lamadon

@tlamadon

Assistant Prof at UChicago

Chicago, IL Katılım Mayıs 2007

224 Takip Edilen407 Takipçiler

Thibaut Lamadon retweetledi

Econometrica@ecmaEditors·15 Kas

We show how and when grouping individuals in a first step, and estimating the model given the groups in a second step, provides consistent estimators even when unobserved heterogeneity is continuous. Check bit.ly/gfe-2021 econometricsociety.org/publications/e…

English

216

Thibaut Lamadon@tlamadon·17 Eki

@thetahat @SouravS19697607 My favs for data work are DataFrames.jl with DataFramesMetas.jl and chain.jl. It’s just amazing, far superior to stata. I now even prefer it to R and python.

English

Thibaut Lamadon@tlamadon·8 Haz

@johannesmboehm @pengzell @paulgp @emanresa82 @raffaelesaggio @JuliaLanguage your words, not mine ;-)!

English

Johannes Boehm@johannesmboehm·8 Haz

@tlamadon @pengzell @paulgp @emanresa82 @raffaelesaggio @JuliaLanguage Thanks @tlamadon ! Yes that's exactly the idea, to lower entry barriers to Julia for Stata folks (and maybe then we can get rid of Stata altogether :) )

English

Per Engzell@pengzell·3 Haz

Question for #EconTwitter: is there a Stata hack for correcting limited mobility bias in AKM decompositions on matched firm-worker data? Or do I have to learn Python?

English

Thibaut Lamadon@tlamadon·8 Haz

@pengzell @paulgp @emanresa82 @raffaelesaggio Following up, has any of the stata people given a try/chance to github.com/jmboehm/Dougla… ? Not sure how many commands people actually use, but barrier to switching to @JuliaLanguage can be quite low! @johannesmboehm, I love the idea!

English

Per Engzell@pengzell·3 Haz

@paulgp @emanresa82 @raffaelesaggio Thanks. It seemed like a general enough thing that I would ask on here first...

English

Thibaut Lamadon retweetledi

Downdetector@downdetector·7 Haz

User reports indicate Showtime Anytime is having problems since 7:57 PM EDT. downdetector.com/status/showtim… RT if you're also having problems #showtimeanytimedown

English

183

123

Thibaut Lamadon@tlamadon·3 Haz

@raffaelesaggio @pengzell Experimental for now, it calls the Python code that can be installed via conda or pip. We need more testing for sure, let us know how it goes if you try!

English

Raffaele Saggio@raffaelesaggio·3 Haz

@pengzell hi! That's great timing, I think Thibaut Lamadon just released a Stata Package that does all you need (BLM,KSS, etc): econpapers.repec.org/software/bocbo…. There is also Julia code to run KSS for the Julia aficionados here: github.com/HighDimensiona…

English

Thibaut Lamadon@tlamadon·3 Haz

@pengzell @LsageSocio We are also putting together a @JuliaLanguage version of the BLM code: github.com/tlamadon/blm.jl, the interface needs polishing but the code works.

English

Per Engzell@pengzell·3 Haz

@LsageSocio I did not know this one. Thank you!

English

Thibaut Lamadon@tlamadon·3 Haz

@arindube @pengzell @paulgp @emanresa82 yes, it is there and working. It wouldn't hurt for us to run a few more test, but the implementation is there. It can even run with multiple cores. We need to add an example to the documentation. We will be doing this asap!

English

Arin Dube@arindube·3 Haz

@tlamadon @pengzell @paulgp @emanresa82 Does it also provide the KSS correction? A quick browse at the help file seemed to have a [TBD] after KSS, so wasn't sure.

English

Thibaut Lamadon@tlamadon·3 Haz

@FlorianOswald Very experimental! But happy to work with people to improve it! Please reach out on the GitHub repo! There is also a command line interface. We even provide the conda compiled packages for easy install.

English

Florian Oswald 🇪🇺 🇺🇦@FlorianOswald·3 Haz

FWIW tlamadon.github.io/pytwoway/doc-s… @tlamadon does it all

Per Engzell@pengzell

@paulgp Should have done that. There's this by @emanresa82 and coauthors nber.org/papers/w27368 with an associated Python package that looks terrific tlamadon.github.io/pytwoway/ ... Just curious why no one seems to have done it in Stata so far

Thibaut Lamadon@tlamadon·3 Haz

@pengzell @paulgp @emanresa82 The main bottle neck in stata/mata is solving a large sparse linear system. Actually it appears MATA doesn’t have great support for sparse matrices as far as I can tell.

English

Per Engzell@pengzell·3 Haz

English

Thibaut Lamadon@tlamadon·11 Nis

@FlorianOswald @HaroldZurcher87 This is really great content to share! Since you are showing it, it does look like a lot of memory allocations and GC, which always makes me curious!Looking forward to seeing the code!

English

Florian Oswald 🇪🇺 🇺🇦@FlorianOswald·8 Nis

Yo @HaroldZurcher87 check it out. In tomorrow's class will mpec the heck out of the good old bus replacement model.

English

Thibaut Lamadon@tlamadon·3 Nis

@AriKatz20 @AlexZevelev @FlorianOswald @_mbauman I have had difficulty getting speed ups with threads, but Distributed has been working really great. I will def check ThreadX and other mentions! Thanks!

English

R-E@akatzzzzz·3 Nis

@AlexZevelev @FlorianOswald @_mbauman see: juliafolds.github.io/data-paralleli… for better than broadcast parallelism. Also, github.com/mcabbott/Tulli… Both more expressive than broadcast, and support GPUs and distributed as well. Easy to make a broadcast threads macro. There's one in strided.jl

English

Florian Oswald 🇪🇺 🇺🇦@FlorianOswald·2 Nis

today's class github.com/floswald/Numer… 1. take 10 million rows of data, 2 FEs 2. FixedEffectModels.jl (1 thread): 12 secs 3. R{lfe} (single thread): 26 secs 4. R{fixest} (4 threads): 7 secs 5. wait...what's that: set `method = :gpu` on FixedEffectModels.jl? => 6 secs

English

Thibaut Lamadon@tlamadon·2 Nis

@FlorianOswald do I see a double_precision = false in the gpu call🤔? I am curious if it matters in this context for the results?

English

Florian Oswald 🇪🇺 🇺🇦@FlorianOswald·2 Nis

where the gpu method probably spends most of the time copying the 10 mio rows of data over to the device.

English

Thibaut Lamadon@tlamadon·19 Mar

@FlorianOswald Why are we still writing the optimizer and the objective function in the same language for cases where function evaluation is the bottle neck? Why not have the optimizer as a web service?

English

Florian Oswald 🇪🇺 🇺🇦@FlorianOswald·18 Mar

Looks like a good algo for people with access to (and skills for) an INTEL fortran compiler. Particularly the async model. A bit like research.cs.wisc.edu/htcondor/ but on your own computers. for #julialang I keep an eye on galacticoptim.sciml.ai/stable/. key: easy access to many algos.

Serdar Ozkan@serdarozkanEN

I have been using this algorithm in my papers with great success. It allows you to search a wide range of parameter space very efficiently.

English

Thibaut Lamadon@tlamadon·10 Mar

@valandau Same thing here, I even looked for alternatives to htop, I like github.com/aristocratos/b… and zenith!

English

Thibaut Lamadon@tlamadon·31 Ara

@gzervas Great stuff! We also have been using pytorch to estimate FE and interacted FE models using GPUs. Auto diff and gpu are remarkable techs!

English

Thibaut Lamadon@tlamadon·25 Eyl

@makotonkjm How about a geek bench 5 link? see how it flies! I just got a ryzen 3950, also nzxt case. Pretty happy with it!

English

Makoto Nakajima@makotonkjm·25 Eyl

Labor Day Weekend Project 2/2 Took longer to set up everything but my new work desktop started working hard doing calibration for me now! (specifically, 20 virtual cores overclocked above 5Ghz).

English

Thibaut Lamadon@tlamadon·29 Ağu

@FlorianOswald I would recommend @Docker as the default packaging for replication. No headaches to rerun! This repo looks extremely well documented though, and that’s an amazing effort! Thanks @tradewartracker!

English

Florian Oswald 🇪🇺 🇺🇦@FlorianOswald·29 Ağu

The replication kit is worth a very close look as there is lots of well written #JuliaLang in github.com/jlperla/PerlaT…

Mike Waugh@tradewartracker

YESSSSSSSSSSSS!!!!!!! This took a long time, but was worth it. I'm very proud of the final product....I'll write more about this later!

English

Keşfet

@SouravS19697607 @johannesmboehm @pengzell @PaulGP @emanresa82 @raffaelesaggio @JuliaLanguage @arindube