Dylan Balla-Elliott

🫡

You could imagine a more picky language throwing an error in those cases and making the user be explicit about what they wanted to do, but stata just runs

To the point about strong typing, I do think that a lot of unintuitive issues in stata come from the fact that it tries to let you run bad code — eg

missing > 5 returns true, not missing

by default, sorting on a non unique variable will break ties randomly (ignoring random seeds!)

I haven’t used it a ton but I’ve really liked the little bit that I’ve done in Julia!

Ah sure. You mean equality checks are dangerous generally because of rounding, whether float or double?

Something like
abs(x-y) < 10^-6

instead of
x==y

Or maybe I’m not following your point.

I should also say that I spent an ungodly amount of time on your website when I was an RA getting used to Stata. It’s a great resource!

There’s a running joke that continuous variables are a modeling fiction because P(X=x) > 0 in any real data set

Apparently this is also true in fake data due to floating point rounding!

Totally!

Just highlighting that the defaults make it easy to accidentally compare floats with doubles.

Yeah this is exactly the float equality issue.

I think Python is actually more transparent in this example, since printing x shows you the float approximation. In the Stata example, tab will display that x = 1.1 (even though that’s only approximately true)

brought to you by a morning of me slowing losing my mind

👻 Spooooky Stata fact

Stata by default uses *float* precision to store data and performs calculations in *double* precision, which means that:

gen x = 1.1
assert x. == 1.1
^^ this is false!!

📈📉 🙃

Good news is bad

Does "Trickle Down" Work?: Economic Development Strategies and Job Chains in Local Labor Markets Persky, Felsenstein, and Carlson explore a new framework for evaluating state and local economic development efforts. They propose a method, referred to as the “job-chains approach,” that they say...

One area of research in labor econ/urban econ in which I think there is a lot of potential for productive future research is in job chain models of local labor markets. #EconSky. For intro to potential of such models, see research.upjohn.org/up_press/25/

this article has a bunch (in political science) and shows that basically the results there do mostly survive the new methods!

ah this is very cool! ty!

#econsky 📉📈 is there an (applied) example of a synthetic control paper that writes down a factor model for the outcome of interest?

Ah well nevertheless

Ah well nevertheless

"Worker Mobility in Production Networks" -- neat paper that brings together employer-employee data and firm-to-firm transactions to demonstrate the role played by production networks in shaping the job search and matching process. #EconSky 📈

🔗 steg.cepr.org/sites/defaul...

unfortunately inspired by a typo in a pset question last year

oh you thought you would just add up all the numbers and divide by N? Did you make sure the distribution even *has* a mean? turns out that thing is super poorly behaved, 100 years of measure theory before you get to open up your laptop again. sorry I don't make the rules.

@economeager.bsky.social had a tweet a while ago about understanding some of OLS on a good day and I think about that a lot

grad school so far:
I hope I can learn how to do good applied work ->
I guess I should learn how IV works so I can do good applied work ->
I guess I should learn how OLS works so I can do good applied work ->
I guess I should learn what a standard error is so I can do good applied work ->
???

I’ll definitely have to include this in the next draft!

Identifying Causal Effects in Information Provision Experiments TSLS can underestimate the effects of beliefs on outcomes. I propose an alternative way to estimate average effects in information provision experiments.

This is super great! I wish I had seen this a day or two earlier and I could have incorporated it into the current draft of this.

A key mechanism in this paper is how the effects of beliefs shape belief formation which then shapes belief updating.

www.dballaelliott.com/papers/info_...

Thanks Vitor!

hi #EconSky! 👋

I’m a third-year econ phd at uchicago studying behavioral development! I just recently returned from scoping work in Nigeria and Uganda investigating trust and family ties in firm decision making.

otherwise, I’m a proud Buffalo native, runner, and fan of electronic and folk music

check out the full paper here: dballaelliott.com/papers/info_iv/

or on arxiv: arxiv.org/abs/2309.11387

In an application, I estimate that the average partial effect is about 40% larger than the TSLS estimate. As theory predicts, this occurs because belief effects are close to zero for the group that updates the most (and thus has the largest TSLS weights).

I show that a Bayesian learning assumption is sufficient* to identify the average effect—instead of a complicated TSLS weighted average—in designs with multiple signals.

(*Bayesian learning is stronger than necessary and as shown in the paper can be relaxed in specific ways.)