Devastating news for the logit supporters.
*US MOVING FORWARD WITH 104% CHINA TARIFFS: OFFICIAL
April 8, 2025 at 6:57 PM
Devastating news for the logit supporters.
new blog post! why do LLMs freak out over the seahorse emoji? i put llama-3.3-70b through its paces with the logit lens to find out, and explain what the logit lens (everyone's favorite underrated interpretability tool) is in the process.
link in reply!
link in reply!
October 5, 2025 at 2:36 PM
new blog post! why do LLMs freak out over the seahorse emoji? i put llama-3.3-70b through its paces with the logit lens to find out, and explain what the logit lens (everyone's favorite underrated interpretability tool) is in the process.
link in reply!
link in reply!
Feeling some appreciation for @lnalborczyk.bsky.social's logit dot plots (lnalborczyk.github.io/blog/2018-01...).
February 5, 2026 at 9:48 PM
Feeling some appreciation for @lnalborczyk.bsky.social's logit dot plots (lnalborczyk.github.io/blog/2018-01...).
Yes, it's a logit subtweet.
April 3, 2025 at 8:52 PM
Yes, it's a logit subtweet.
logit, probit, cauchit, cloglog, identity, log
what are the Six Genders? wrong answers only
May 15, 2026 at 6:38 PM
logit, probit, cauchit, cloglog, identity, log
LPM vs logit, round 1,520,910,462
Question for the quants: Is it justified to have a general preference against logit models? The interpretation problems presented by log-odds ratios basically make me think that logit should only be used after there's a formal justification re: why linear models will be mis specified
April 5, 2026 at 7:56 PM
LPM vs logit, round 1,520,910,462
Like logit wars, I guess Blue Sky has arrived if we’re now having Stata vs R wars.
This is why our department has mostly switched to giving our grad students code examples in R. They will mostly be at places with less resources than us, won't have research budgets to buy Stata licenses, and more likely to collaborate with others in R. But Stata endures for four main reasons:
December 14, 2024 at 4:18 PM
Like logit wars, I guess Blue Sky has arrived if we’re now having Stata vs R wars.
Sociology: you can’t compare logit or probit coefficients across models
Psychology:
Psychology:
3. I then found out that the author mixed suicide RATES (per 100,000) with suicide RATE RATIOS either standardized to 1 or to 100. See my comments to the figure. I confirmed this by looking at several individual studies. Of course this doesn't make sense at all.
August 18, 2025 at 7:12 AM
Sociology: you can’t compare logit or probit coefficients across models
Psychology:
Psychology:
November 18, 2024 at 10:21 PM
Just to make this into Econ Twitter, LPM is probably better than logit/probit for most cases 😉
One thing that is usually not communicated to students enough is the cost of learning R/Python increases as you become better at Stata.
I'm not an R person, but I totally agree with this point: we're doing students a disservice by teaching them stata.
Even worse if you're teaching at a public university: using public money to create consumers for a private company.
Even worse if you're teaching at a public university: using public money to create consumers for a private company.
December 14, 2024 at 2:44 PM
Just to make this into Econ Twitter, LPM is probably better than logit/probit for most cases 😉
You know what’s the real disappointment here? We never went all the way. We never had a logit vs probit showdown.
I miss old academic twitter.
New academic Twitter and Bsky just aren't the same without the logit/probit vs. LPM and Stata vs. R throwdowns.
New academic Twitter and Bsky just aren't the same without the logit/probit vs. LPM and Stata vs. R throwdowns.
December 13, 2024 at 6:24 PM
You know what’s the real disappointment here? We never went all the way. We never had a logit vs probit showdown.
Related: Even in complicated DiD designs, imputation with logit and pooled estimation with logit are identical. Not so for probit. So logit is a bit more robust. Maybe a minor difference in practice, but still there.
Probit often easier to work with when instruments are needed.
Probit often easier to work with when instruments are needed.
In completely randomized experiments, avg marginal effects from logit MLE or OLS (with pre-treatment covariates) are consistent for the avg treatment effect even if the model's wrong. Not true of probit MLE. This old tweet links to a helpful thread by @jmwooldridge.bsky.social
x.com/linstonwin/s...
x.com/linstonwin/s...
x.com
x.com
December 16, 2024 at 12:53 PM
Related: Even in complicated DiD designs, imputation with logit and pooled estimation with logit are identical. Not so for probit. So logit is a bit more robust. Maybe a minor difference in practice, but still there.
Probit often easier to work with when instruments are needed.
Probit often easier to work with when instruments are needed.
in a compromise solution, I believe we should make logit regression illegal
March 11, 2026 at 9:40 PM
in a compromise solution, I believe we should make logit regression illegal
Tired: logit model as "machine learning"
Wired: ctrl+F as "AI"
Inspired: cuts for you, govt contracts for me as "efficiency"
Wired: ctrl+F as "AI"
Inspired: cuts for you, govt contracts for me as "efficiency"
honest question: did they just ctrl+f for "gay" bsky.app/profile/brad...
Brad Heath (@bradheath.bsky.social)
The military is plans to scrub photos of the Enola Gay, one of the bombers that dropped an atomic bomb on Japan, from its website as part of its purge of diversity content.
It's among more than 26,0...
bsky.app
March 7, 2025 at 3:42 AM
Tired: logit model as "machine learning"
Wired: ctrl+F as "AI"
Inspired: cuts for you, govt contracts for me as "efficiency"
Wired: ctrl+F as "AI"
Inspired: cuts for you, govt contracts for me as "efficiency"
I know it looks like a Logit but it's actually machine learning
March 13, 2025 at 11:55 PM
I know it looks like a Logit but it's actually machine learning
This whole thing is so full of researcher degrees of freedom I'm going to go insane. Why does he bin everything into less than or greater than 50? Why switch from a linear probability model to a logit? Why?
December 16, 2025 at 1:26 AM
This whole thing is so full of researcher degrees of freedom I'm going to go insane. Why does he bin everything into less than or greater than 50? Why switch from a linear probability model to a logit? Why?
🚨 Blog post:
Lost in Transformation: The Horror and Wonder of Logit
I use the analogy of 🗺️↔️🌏 to clarify why logit is such a good model for binary outcomes.
tl; dr: OLS works on flat maps; logit works on globes.
statsky polisky econsky academicsky
www.robertkubinec.com/post/flat_ea...
Lost in Transformation: The Horror and Wonder of Logit
I use the analogy of 🗺️↔️🌏 to clarify why logit is such a good model for binary outcomes.
tl; dr: OLS works on flat maps; logit works on globes.
statsky polisky econsky academicsky
www.robertkubinec.com/post/flat_ea...
Lost in Transformation: The Horror and Wonder of Logit | Robert Kubinec
Perhaps no other subject in applied statistics and machine learning has caused people as much trouble as the humble logit model. Most people who learn either subject start off with some form of linear...
www.robertkubinec.com
April 10, 2024 at 10:02 AM
🚨 Blog post:
Lost in Transformation: The Horror and Wonder of Logit
I use the analogy of 🗺️↔️🌏 to clarify why logit is such a good model for binary outcomes.
tl; dr: OLS works on flat maps; logit works on globes.
statsky polisky econsky academicsky
www.robertkubinec.com/post/flat_ea...
Lost in Transformation: The Horror and Wonder of Logit
I use the analogy of 🗺️↔️🌏 to clarify why logit is such a good model for binary outcomes.
tl; dr: OLS works on flat maps; logit works on globes.
statsky polisky econsky academicsky
www.robertkubinec.com/post/flat_ea...
Also, check out the second post in the series, showing the equivalence between conditional logit, multinomial logit, and log-linear models of count data for typical model specifications. benjaminfjarvis.com/blog/clogit-...
Log-Linear vs. Multinomial vs. Conditional Logit Models of Assortative Mating – Benjamin F. Jarvis
The second note in the series. Here I estimate and compare parameter estimates from conditional logit, multinomial logit, and loglinear models of two-way contingency tables (or data that can be reduce...
benjaminfjarvis.com
September 24, 2026 at 1:49 PM
Also, check out the second post in the series, showing the equivalence between conditional logit, multinomial logit, and log-linear models of count data for typical model specifications. benjaminfjarvis.com/blog/clogit-...
Fun fact: you can estimate a conditional logit by running a fixed effects Poisson. They yield the same estimates and standard errors - they maximize the same likelihood. Poisson regression forever!
October 15, 2025 at 6:27 PM
Fun fact: you can estimate a conditional logit by running a fixed effects Poisson. They yield the same estimates and standard errors - they maximize the same likelihood. Poisson regression forever!
Data science becomes like a thousand times less intimidating coming from an academic background when you realize they're just building models and calling it something different. "I built a classifier" could just mean they ran logit. OLS can be a "classifier"
September 25, 2025 at 1:40 AM
Data science becomes like a thousand times less intimidating coming from an academic background when you realize they're just building models and calling it something different. "I built a classifier" could just mean they ran logit. OLS can be a "classifier"
The LPM/logit wars can finally be put to bed. Logit wins because it's AI. Checkmate.
January 14, 2026 at 11:48 PM
The LPM/logit wars can finally be put to bed. Logit wins because it's AI. Checkmate.
I get the impression that polisci is now pretty sold on LPM over logit. What's the go to citation making the case for this?
December 3, 2024 at 4:35 PM
I get the impression that polisci is now pretty sold on LPM over logit. What's the go to citation making the case for this?
Stupid logit, always ruining everything...
November 18, 2025 at 2:17 PM
Stupid logit, always ruining everything...
This kind of thing does not fill me with confidence that British think tanks and politicians aren’t going to be totally snowed by AI/ML advocates selling them a logit model for £10m.
I know it's petty, but describing things as machine learning when they existed for a long time as part of statistics is a red flag for me. k-means is machine learning, regression is machine learning, t-tests are machine learning, everything is machine learning.
Strikes me this bit from the TBI's new report is redolent of a wider, ongoing tendency to de-emphasise (and therefore normalise) the essential ideology of the far right, but hey-ho. (Source: institute.global/insights/pol...)
January 31, 2025 at 8:11 AM
This kind of thing does not fill me with confidence that British think tanks and politicians aren’t going to be totally snowed by AI/ML advocates selling them a logit model for £10m.
Reposting this in Bluesky, my recent post describing why logit is widely misunderstood:
www.robertkubinec.com/post/flat_ea...
www.robertkubinec.com/post/flat_ea...
November 13, 2024 at 11:56 PM
Reposting this in Bluesky, my recent post describing why logit is widely misunderstood:
www.robertkubinec.com/post/flat_ea...
www.robertkubinec.com/post/flat_ea...