@IrvineWelsh We can now append felonious fanta faced falangist ...
English
Roger Koenker
187 posts
Roger Koenker
@rkoenker
Honorary Professor of unmeaning.
London Katılım Mart 2009
94 Takip Edilen705 Takipçiler
Is there a sentient being left on the planet who actually believes a single word this Fanta Faced Falangist Fuckpuppy spouts? twitter.com/realdonaldtrum…
Donald J. Trump@realDonaldTrump
Only 109 people out of 325,000 were detained and held for questioning. Big problems at airports were caused by Delta computer outage,.....
English
@paulnovosad I still recall Cecilia presenting this paper in a seminar at UIUC before it was published, and I still regard it as one of the most careful and original applied micro talks I've ever encountered.
English
Roger Koenker retweetledi
Hi #EconTwitter!📈
Interested in Empirical Bayes methods and their applications in #economics?
Don't miss the next Chamberlain Seminar tutorial (this Friday, November 17, at 9am PT / noon ET / 5pm London) titled
Empirical Bayes: Methods and Applications
by @rkoenker (@ucl ) & @PpM_O (@Cornell), with moderation by Alberto Abadie (@MIT). ⭐️
Cool stuff!
Link to registration: stanford.zoom.us/meeting/regist…
English
@predict_addict @ChristophMolnar This is a truly, madly deeply irritating view. Is Stein shrinkage outmoded because it arrived in the 50's?
QR is a general way to estimate conditional quantile functions, comes with asymptotic inference and underlies a central conformal inference paper.arxiv.org/abs/1905.03222
English
Furthermore, relying on quantile regression can mislead individuals into believing they are adequately quantifying uncertainty when, in reality, they aren't. I've witnessed actual companies invest over a year in building extensive forecasting systems using deep learning and quantile regression, only to see these systems fail spectacularly in production. Such failures can severely harm customer trust and result in significant financial losses.
Quantile regression, a technology from 1978, is an outdated approach for uncertainty quantification, especially when superior alternatives exist today.
English
xgboost 2.0 is out!
What stands out: xgboost can now natively handle many additional prediction tasks:
- learning to rank
- quantile regression
- multi-target regression
- multi-label classification
- multi-class classification
xgboost becomes more and more of an all-rounder
English
@NialFriel Indeed, and I see it as quite a challenge to those who adhere to the medieval notion of science as a bunch of alchemists toiling alone in their garrets hoping to strike it rich with the aid of the patent system. Or their hegemonic corporate counterparts.
English
arxiv.org/abs/2310.00865
This is a really interesting paper by David Donoho on what has driven the rapid explosion of AI research and how this has been (incorrectly) perceived by the public.
Spoiler: The maturity of Data Science has been the driver of this change! 1/4
English
@SolomonKurz Haavelmo thought that all our "little" misspecification errors satisfied a Lindeberg condition.
English
When using OLS regression, from where do we get the assumption the residuals are normally distributed? You don't need that assumption to minimize the sum of the squared residuals, nor to use the CLT to compute the SE's for the beta coefficients. So what's its origins?
English
@predict_addict @selcukorkmaz Practical decision making is a big subject, only a small part of which is covered by prediction intervals.
English
Also I would take an issue with statements like this “However,thereareseveralremainingissues,not theleastofwhichiswhethertheperformanceguaranteesofferedbythemethodarerelevantto practicaldecisionmaking”. Of course performance guarantees are important, it has been shows that some classical models declared 95% percent coverage whilst often delivering numbers as low as 20%.
English
A Simple Guide on Quantile Regression
🧵1/ Introduction to Quantile Regression 📈
Ever noticed how average predictions (like mean) might not capture the entire story of your data? Enter Quantile Regression (QR)! Instead of focusing just on the mean, QR looks at various quantiles (percentiles) of the response variable.
2/ Traditional vs. Quantile Regression 📊
Traditional linear regression predicts the mean of the dependent variable. But what if we're interested in, say, the median? Or the 90th percentile? QR allows us to model these specific quantiles, providing a fuller picture of the data's distribution.
3/ Why use Quantile Regression? 🤔
• To understand the relationship at various points (quantiles) of your dependent variable.
• Highly robust to outliers.
• Helpful when the residuals of a linear model aren’t homoscedastic (i.e., they have non-constant variance).
4/ How does it work? 🛠️
QR minimizes the sum of weighted absolute residuals, unlike least squares regression which minimizes squared residuals. By changing the weights, we target different quantiles.
5/ When to use Quantile Regression? 📅
• When you suspect heteroscedasticity.
• To analyze the impact of variables at different parts of the distribution.
• When interested in high or low extremes (e.g., what factors influence the top 10% of incomes).
6/ Visualization Power 🌈
Plotting several quantile regressions together can give a more holistic view of the data relationship. For instance, seeing how the effect of education on income changes across the income distribution.
7/ Limitations 🚫
• Can be computationally intensive for large datasets.
• Interpretation might be less intuitive than mean-focused methods.
8/ In Conclusion 🎓
Quantile Regression offers a versatile tool to understand relationships in your data that go beyond the average. It shines a light on the entire distribution, allowing for richer insights.
9/ Further Reading 📚
For those keen on diving deeper, many statistical packages, like R's quantreg, offer tools to implement and visualize QR. #Rstats
10/ Liked this thread? 🌟
Feel free to like, retweet, and share your experiences with Quantile Regression below! #Statistics #DataAnalysis #DataScience #QuantileRegression
English
@predict_addict @selcukorkmaz Actually, to be fair, in this example, conformalization doesn't make a perceptible difference to what one can see with total variation penalized ordinary quantile regression.
econ.uiuc.edu/~roger/researc…
English
@selcukorkmaz Much better Conformalized Quantile Regression
“How to predict quantiles in a more intelligent way (or ‘Bye-bye quantile regression, hello Conformalized Quantile Regression’)”
@valeman/how-to-predict-quantiles-in-a-more-intelligent-way-or-bye-bye-quantile-regression-hello-24a65e4c50f" target="_blank" rel="nofollow noopener">medium.com/@valeman/how-t…
English
@KevinDenny @selcukorkmaz It would be except that regression is inherently conditional, so unconditional quantile regression is a bit of an oxymoron.
English
@selcukorkmaz It’s important to distinguish between conditional and unconditional QR.
English
@HOS_ASA It is important to read beyond the title of this!
tandfonline.com/doi/abs/10.119…
English
#OTD 1941 Stephen Jay Gould b (d 20 May 2002) 🇺🇸paleontologist, evolutionary biologist, science writer. Best known (to statisticians) for his classic article on statistical thinking ‘The Median Isn't the Message’ people.umass.edu/biep540w/pdf/S…
English
@stephensenn @yudapearl This is admirably clear. I can't resist mentioning the connection to one of my favorite ancient papers:
jstor.org/stable/2332424
English
@yudapearl But statistics has a formalism for designed experiments that can distinguish between the two versions of Lord’s Paradox (Book of Why & primer) that it seems causal methods struggle with. linkedin.com/pulse/lauding-…?
English
This quote by De Morgan should be studied in greater depth: "Every science that has thriven has thriven upon its own symbols: logic, the only science which is admitted to have made no improvements in century after century, is the only one which has grown no symbols." (Augustus De Morgan, 1864). I usually quote the first part (#Bookofwhy) but the second part is as profound: Behold how formal logic was held back for centuries, and logicians are no stupid, just for lack of notation/calculus. For statisticians, this should send a somber soul-searching message: Isn't it possible that statistics too was held back by lack of notation, and is now due for revival and resurgence? Just a thought.
Judea Pearl@yudapearl
Beg to differ. It's equally likely that the proper mathematics has not been developed for that field. Take logic, which got stuck for two thousands years in Aristotle syllogisms, then burst forward with the advent of Boolean Algebra (1847). "Every science that has thriven has thriven upon its own symbols: logic, the only science which is admitted to have made no improvements in century after century, is the only one which has grown no symbols." (Augustus De Morgan, 1863). I wonder where statistics stands in its openness to new mathematics?
English
@JohnMullahy I wrote something similar in R for students presenting pieces of problem sets. Of course this sort of thing is subject to: xkcd.com/1205/
English
New typo almost introduced (by me) in 8e of my intro book: "Derive the attention bias in the OLS estimator ...."
Maybe I should leave it as a kind of Easter egg.
English
@f2harrell @vandy_biostat My mantra: Every parameter would like to grow up to be a distribution.
English
#Statistics thought of the day: What do fear of the proportional hazards/odds assumption, testing for normality, testing for equal variance, and use of restricted mean survival time have in common? Fear of devoting more than one parameter to treatment. @vandy_biostat
English
@and_joy_ All of Alfred Marshall qualifies, and Joan Robinson had a knack for writing mathematical arguments without mathematics thereby making them unnecessarily obscure.
English
What is your favorite paper with the highest insight to formalism/math ratio?
Not talking about hand-wavy insight/conjectures, but economic ideas that flow like proofs written in prose.
English
@jmwooldridge Rao visited UIUC in 1983. There was a lunch in a dive called "The Hip Pocket", someone asked him about the Cramer-Rao inequality, and he replied that "it was much less deep than a random line from Ramanujan's notebooks." Respect.
English
I missed the announcement of CR Rao’s passing! A wonderful man and such an impressive life. I got to meet him when he came to MSU several years back. I have a chapter in a book in his honor, and he graciously signed it. I probably acted like a super fan boy.
Giuseppe Cavaliere@CavaliereGiu
CR Rao was one of the greatest statisticians. He was awarded the National Medal of Science — a very exclusive honor bestowed by the President of the United States — for math. Do you know of any other US statistician (or prob) who received this medal?
English
So many brilliant ideas: Two of my favorites:
Hotelling, H. (1939). Tubes and spheres in n-space and a class of statistical problems. Am J of Math, and Hotelling, H. (1933) Review of Secrist, JASA, 28, 463-4. Don't forget marketing ice cream.
ASA History of Statistics Special Interest Group@HOS_ASA
In Aug 1931 Harold Hotelling (1895-1973) 🇺🇸published ‘The Generalization of Student's Ratio’. He generalized Student's t-test to the simultaneous test of hypotheses of differences between means for multiple joint normally distributed variables. 1/2
English
"When you prove a theorem, then you get a solid basement once and forever.”
English
@predict_addict @DataSciFact Sure, but Thomas Bayes younger brother, Empirical Bayes was ok, even Neyman like him.
English
@DataSciFact Indeed, the greatest statistical minds like Fischer and Neyman held Bayesian methods in disdain.
English
R. A. Fisher was the first to use the term 'Bayesian,' and he didn't mean it as a compliment.
English