🚨New Paper Alert 🚨
How do you train a model to make forecasts about a totally new domain?
How about one with no data to train on?
Sounds impossible.
We propose and tested a way and it worked, well, well! 🎯
Okay so this is wild. I'm using this ticket machine in Japan and it malfunctions and doesn't give me my change.
Suddenly, a panel in the wall opens in the wall and a guy appears and tells me to hold on and then gives me my change.
To do something about the high price of gas, Germany has introduced a €9 ticket that lets you travel all around the country.
Basically free transit for an entire month.
I'm here for 3 days and bought one.
@Vamleaf
When he told me to hold on I realized I had a rare opportunity to capture something that people would doubt if I didn't. Gave me time to get my phone ready.
Suppose f is a real-valued function defined on the interval [a, b] and integrable in the sense of Lebesgue. Now suppose I was lying about that. Do you feel betrayed?
One person I feel who is insufficiently appreciated is Hadley Wickham. The results of his work are everywhere, every journal and conference I know. He changed the history of science.
At the breakfast bar and wondering if having everybody touch the same pair of tongs in the same place is proven to be better than having everybody simply pick up their own piece of bread
Econ academics will spend eight days polishing a 20 page reply-to-reviewers letter and then go to the FTC and present something they slapped together on the train
This is said without exaggeration
If you see a graph like this, I don't think you should conclude people are "really bad" at estimating something (like where they sit in the income distribution or the size of groups in America), unless you truly believe they're "really bad" at estimating everything.
1/4
Covariance can be visually represented with the areas of rectangles formed by the origin and each point.
From Wagenmakers,Godmann, de Ruiter, van Doorn. Visualizing the Equation for the Sample Correlation Coefficient
AI writes the paper. AI reviews the paper. AI rejects the paper. AI sends to paper to another AI. New AI says revise and resubmit. AI revises and resubmits. AI accepts paper. AI tweets about the paper.
1/
Want to help your students write impactful articles? Share this tip.
Round your darn numbers!
We've got empirical evidence by not rounding:
-Readers are much less likely to recall (even roughly) your findings
-Readers are much less able to apply your findings
1/
Kids are hard to understand
Our 8 year old:
- made a fake German passport
- made herself unemployed
- made the city of issue Bielefeld
We have no idea why
I know a person who runs lots of field experiments. Literally prize winning stuff. Their estimate of how likely it is for a field experiment to happen successfully once a partner firm or organization agrees:
10%
(Success here means actually carried out, correctly)
"Under attack for decades, rational choice theory finally died in 2020 when a majority of economists admitted it could not explain their own use of STATA."
Microsoft Research New York City (MSR-NYC) and Microsoft Research New England (MSR-NE in Cambridge MA) will do calls for postdocs in:
Computational Social Science
Economics and
Econ-CS (Economics+Computer Science)
I will tweet and toot these announcements as they come.
P-hacking can be time consuming and tedious. Proud to open source this efficient base R routine.
- Runs in < 500 nanoseconds (per microbenchmark)
- Easily parallelizable
- Output preserves monotonicity and linear scalability in its defined domain
🎉🎉🎉
The Computational Social Science group at
@MSFTResearch
New York City is now accepting applications for PhD student interns for 2023!
Internships are:
- Paid!
- 12 weeks!
- In person in NYC! (w/ up to 50% work from home!)
1/4
Please note that there is a huge debate whether the first version is an error given conversational norms. The second one is more clearly an error and reflects current difficulties in getting these things to do mathematical reasoning.
Imagine you:
Take a random sample of 100 observations from a standard normal distribution (mean 0, standard deviation 1). Take another. Do a t-test on the two samples and record the p value. Repeat the above 10,000 times.
A histogram of the 10,000 p values will look like a
Are there readings or decision trees to help students learn how to choose a mixed-effects (random effects) model specification?
There's much written on syntax, less on the principles.
Something with examples of both (and newer than Gelman & Hill) would be great!
If you're interested in this, read this paper by Miller and Sanjurjo
I once interviewed Miller on how they figured something out that people studying the hot hand had missed for decades
A Harvard economics student Wesley Wang spent 2+ years making a short film in high school and posted it online. Darren Aronofsky saw it & emailed him. After a bidding war, TriStar Pictures bought the rights. Wang will adapt it to film to be produced by Aronofsky’s Protozoa firm.
I wrote up a gotcha that can get you when designing experiments. It's in the intro of the 2022 Behavioral Economics Guide.
Reviewers have described it as "outstanding" and "cracking."
I would be happy if you read it!
Direct link (PDF pages 7-22):
🚨 RESEARCH ASSISTANT POSITION AT MICROSOFT RESEARCH NYC 🚨
Ideal for those who wish to go to PhD programs in Computer Science or Computational Social Science
Strong front-end and R programming skills required
RESEARCH ASSISTANT POSITION AT MICROSOFT RESEARCH NYC
Ideal for those who wish to go to PhD programs in computational social science
Front-end programming skills required
No guarantees but our last 5 RAs *all* got into great grad schools!
It's funny to see praise for these university endowment portfolio managers who did *checks notes* the same as or worse than the S&P 500 while charging hefty fees
it is really next level how everything druckenmiller touches turns to gold. bowdoin endowment, whose investment committee he's chaired for ~20 yrs, retains the
#1
spot on the league tables
Screen time is associated with lower well-being about as much as regularly eating potatoes is. Putting "variance explained" into perspective from
h/t
@jakehofman
Microsoft Research NYC has 12 week internships in Computational Social Science (my group). Goal is to crush a project that we (incl you) find interesting and that is also interesting to academic and general audiences.
@dggoldst
@jakehofman
@DavMicRot
Watching the upper right cell makes me feel guilty about how I've been fooled in the past, angry about the time I've wasted in misguided talks, and relieved that those days are over.
🚨Microsoft PhD internship alert🚨
PhD students: The Microsoft Research NYC Computational Social Science group is really good.
Apply to do a 12 week project of mutual interest with
@jakehofman
@davmicrot
&
@dggoldst
Winter / spring starts encouraged!
I predict that LLMs will be used to flood the internet with pages containing plausible-sounding but worthless answers to vast numbers of search queries. Why? For traffic.
Search engines will spend a lot fighting this.
People will waste a lot of time reading nonanswers.
1/2
You know all those surveys showing how people overestimate the percentage of immigrants in their country?
This working paper shows that if you drop people who are clearly rounding (5, 10, 15, ...) you end up with highly precise estimates!
does anyone have a good explanation for why so many distinguished scientists become crackpots by middle age? is it that thinking hard for a long time makes you go crazy?
I was once designing a study with an econ colleague
-Econ: Aren't we worried about selection there?
-Me: No, because randomization
-Econ: I'm so used to worrying about that I sometimes forget to not worry about that
@uri_sohn
BTW I also bought public transit tickets the time I visited you in Chile. I remember being on a packed bus thinking this is likely the highest density of Chileans I will ever experience.
Statistical significance and even sign can be a distraction when effect sizes are tiny. Effect sizes matter. People's understanding of effect size matters.
You will see a bunch of Meta / Facebook papers drop today and soon after
Hats off to the academic researchers, but also to the internal Meta researchers
The whole idea of doing collaborations with academics was theirs
They invested years of tireless effort to get to today
Behavioral science is over a century old, spans many fields, and shaped modern statistics, experimentation, and machine learning
I keep hearing people dismiss all of behavioral science because of recent fraud in one subfield
Are most people really loss-averse? Not really!
Around 50% of the US population is loss-tolerant and many people accept negative expected-value gambles.
Fascinating new paper by Chapman,
@snowberg
, Wang &
@CFCamerer
that completely changed my priors:
Here's a paper that explains the difference between standard errors and standard deviations:
Here's our paper, which shows that experts (doctors, professors, data scientists) mix them up:
Microsoft Research NYC is hiring a postdoc in "Machine Learning + X," where X is in the set of: AI for Systems, Computational Social Science, Econ-CS, FATE.
Deadline: Dec 15th, 2021
RTs appreciated by the person who gets the job because of you.
IF you are a CS major or skilled front-end programmer
AND you are looking to go to grad school in CS, Psych, Econ, or Business
THEN you should take a year and work with us
(No guarantees but our last 3 RAs got into MIT & Stanford)
This is a pan and plate scraper. I use it daily.
Lasts forever. Hygienic. Dishwasher safe. Won't scratch. Big. Feels good in the hand. Makes sponges mostly irrelevant.
I gave one to my parents, put one in the tool box, drilled a hole in one and hung it in the workshop.
$1.82
NYC undergrads: Want to learn data science this summer from Microsoft researchers?
Each student gets a $5k stipend & a laptop to keep.
As we are committed to diversity in computer science, we strongly encourage women, minorities, and individuals with disabilities to apply.
Help me understand
Presumably every person involved at every step of creating and publishing this map in the Washington Post is a pro specializing in data communication
What chain of failures does it take for a map that wouldn't even get a C in a data viz class to go live?
I regret the time I spend scrolling around on Twitter so I'm quitting for the rest of the year.
Account will stay up and blog autoposts will continue.
Apologies in advance for not giving the ❤️s my counterfactual self would have given.
Why are people error-prone according to some psychologists and intuitive statisticians according to others?
Analysis of 604 studies finds experiments with descriptive scenarios are associated with the error-prone conclusion, experiments involving learning suggest otherwise.
Are our statistical intuitions sound and fit for purpose, or hopelessly inadequate? These conflicting views may be traceable to a shift in experimental methods triggered by the heuristics-and-biases program, paper argues:
HT
@dggoldst
cc
@DocGrawitch
We’re hiring a research assistant in computational social science at Microsoft Research NYC. Good transition path to grad school. If you have any talented - programmer undergrads finishing up or wish to spread the word, we’d love it.
🚨Happy to announce a new paper!
An illusion of predictability in scientific results: Even experts confuse inferential uncertainty and outcome variability
Open access:
The new Cohen's d effect size categories are here:
d < .2: Get out of here
d = .2: Nothing to write home about
.2 < d < .5: Not too shabby
d = .5: Now we're talking
.5 < d < .8: Whoomp there it is
.8 <= h: Get outta here
Movie plot: Assistant professor pays the mafia to randomly disrupt things for the purpose of doing causal inference. Quickly loses control of the situation.
@andreamatranga
Science trivia: The social psychology completeness theorem states that there is a theory from social psychology consistent with every hypothesis in every science.
🚨Postdoc Alert 🚨
The Computational Social Science group at Microsoft Research NYC is seeking a post doc researcher!
Join
@jakehofman
@DavMicRot
and
@dggoldst
for a spell and take your computational social science research game to the next level!