jiafengkevinchen.bsky.social
postdoc at siepr | assistant professor (‘25) of economics at stanford
jiafengkevinchen.github.io
75 posts
2,153 followers
436 following
Getting Started
Active Commenter
comment in response to
post
Ha, have a little stage fright about sending to you this first draft, but obviously very curious about your reaction!
thanks for two great presentations (clearly I learned a lot!)
comment in response to
post
Looking forward!
comment in response to
post
But @jannspiess.bsky.social would know :)
comment in response to
post
This sounds like the exact setting of coussens and spiess (?)
arxiv.org/abs/2108.03726
comment in response to
post
If I understand correctly, it’s doing the same conceptual
thing as rdhonest but imposing that CATE(x) is linear. If so, then probably can extend to discrete running variables (as long as CATE(x) being linear in x still makes sense)
comment in response to
post
One could do a cross-fit version of this where you flip the roles of the train-test split. (This is basically SSIV, but hopefully some machine learning method learns E[treatment | instrument] well)
comment in response to
post
Ah, here's a cheap thing. If you're willing to split sample into (train, test), you could use the training set to compute *whatever* estimate of E[treatment | instrument] you'd like. Call what you get \hat f(z). Then on the test set, get the predictions \hat f(z) and 2SLS w/ \hat f(z)
comment in response to
post
Believe 2sls with multiple instruments is equivalent to 2sls using the first-stage fitted value as an instrument. Even if z2 is just noise (but relatively low dimensional), that’s not going to change the fitted value by much if n is reasonably large. I would just include both
comment in response to
post
I also wondered about this…
“The multiset of vectors x_m, m \in M, is linearly independent”?
comment in response to
post
“October 19, 1989” lol
comment in response to
post
1. consider the hilbert space of assets
2. the map that sends an asset to "its price in cash" is a rank 1 operator on this space
3. consider the singular value decomposition of this operator
4. admire the right singular vector
5. ???
6. profit
comment in response to
post
so if the EO were to stand, a pregnant woman legally permitted to stay/work in the US would need to exit the US, give birth, then apply for a dependent visa for her newborn, right?
comment in response to
post
It's also not clear—were it the policy—whether there is any legal channel to obtain a visa for someone's pending child?
comment in response to
post
err, i just learned about this - doesn't look like a mistranslation to me
maybe "stale and dry" for 干巴？
comment in response to
post
my inner frequentist
comment in response to
post
en.wikipedia.org/wiki/David_S...
I think one can argue this (collab w/ professionals) counts as an amateur breakthrough in math
comment in response to
post
www.quantamagazine.org/teenager-sol...
does this count as amateur?
comment in response to
post
Dumb Q: is this the complete class theorem?
comment in response to
post
I don’t think this contradicts? Do and talk about actually popular things and avoid taking unpopular positions, regardless of what the Groups think
comment in response to
post
average: 51-70%
comment in response to
post
you lost me at "a proof of the ABC conjecture would be written so that [...] people could understand it"
comment in response to
post
maybe it's a good critique of the credibility revolution
comment in response to
post
www.ndtv.com/world-news/f....
comment in response to
post
If you told fisher that one day all of science + FDA uses a rule he just made up cuz 1/20 sounds unlikely enough, I’m not sure he would think the statistics project has failed because it’s irrelevant
comment in response to
post
Wait, but like a normative discipline is by definition distinct from describing how people actually do things? Like ofc what we ought to do doesn’t line up always with what we actually do
(I’m saying this as someone who really thinks methodology should meet practitioners where they are)
comment in response to
post
From reading Erich lehmann’s autobiography, it seemed that a lot of statistics was done because problems encountered during WWII (eg german tank, good-turing estimation, etc.). To your point idk if any of that made a difference in the war, but seemed that the generals cared enough to invest
comment in response to
post
Yup, but I think this points to, like, at least someone in the market appears willing to pay for stats expertise, and not just for academic consumption
comment in response to
post
It’s possible it’s a Stanford thing, but I’d say a nontrivial portion of research at stats, OIT or econ cater very much to folks in industry, often with collaboration (they also seem to be happy to hire PhD students for interns and FT jobs)
comment in response to
post
No the people thinking about AB testing, how to AB test in complex environments, etc.
comment in response to
post
Like the Stanford stats to industry pipeline is still fairly robust. To what extent it’s just signaling on underlying skill i don’t know
comment in response to
post
I might be wrong on this, but my understanding is a fair amount of people doing this have backgrounds (PhDs) in econ and stats
comment in response to
post
I mean, thanks to statistics, we know basically exactly what size the reply-tweet button is to maximize engagement
Isn’t this full circle?
comment in response to
post
i just think it’s a beautiful testament to human curiosity that we know the exact extent to which the evidence supports whether an English lady can taste the difference between milk in tea and tea in milk
comment in response to
post
cut to the chase, alz, what have you done to arpit?
comment in response to
post
I see. Yes, eqm has always been the white whale - Could do (P, D) = g(exogenous) where g takes care of the solving-for-eqm part I guess.
comment in response to
post
chaotic neutral
comment in response to
post
you-know-im-something-of-a-causal-inference-guy-myself.jpg
comment in response to
post
probably true that translating IO to Rubin-style framework is more burden than it's worth - my prior is that it's feasible at least in principle though
comment in response to
post
I get that, but you could say that in the standard Imbens-Angrist IV setup, the endo treatment D = D(z) is also simultaneously an outcome and a treatment?
comment in response to
post
Can I think of D(p) = D(p,u) as the outcome. if I completely randomized p I could learn D(p). Failing that, I’m making more assumptions on either how p is generated or how D is generated or both?
comment in response to
post
I see, I think we have a disagreement in degree but not in kind?
w/ continuous treatments, sometimes people assume "linear effects, but allowed to be heterogeneous, in x"
y_i(x) = beta_i x + eps
and talk about assumptions on x. I would call that (mostly) design-based. obv there's modeling too
comment in response to
post
Seems general enough outside of RCM? (though, what’s an example that can’t be forced into RCM notation?)