macartan.bsky.social
WZB Berlin, Humboldt University, TCD
https://macartan.github.io/
Politics, conflict, inequality, political economy of development, causal inference
68 posts
3,482 followers
308 following
Prolific Poster
Active Commenter
comment in response to
post
throughout, 🇩🇪 has not been having the open, uncomfortable, conversations it needs to have about its role and responsibility
universities have failed to provide a space for these conversations
when they have, political actors have been shutting them down, time and again
it's all beyond shameful
comment in response to
post
Everyone worries about AI and democracy, but AI is just an extension of capital. THE fundamental social problems are power and fairness. What we can do how easily with technologies (e.g. communications like AI, transport, and military), alters the abilities and therefore obligations of us all.
comment in response to
post
99% this! I find it very hard to read arguments about the right estimator when it's not clear what the estimand is. Maybe it is a version of a BLP? Maybe its a large collection of CATEs? Or an average of local differences in CATEs? These different approaches fare differently for different estimands
comment in response to
post
1. Dt. Medien haben insgesamt in ihrer Funktion als sog. vierte Gewalt versagt. Es wäre die Aufgabe der Medien gewesen, der Bundesregierung täglich aufzuzeigen, welchen Horror sie in Gaza unterstützt und dass sie damit gegen internationales Recht + gegen die dt Verfassung verstößt. 3/n
comment in response to
post
[edited] Fair call! I edited to clarify this is about the Y on X regression, not specifically least squares
Y on X targets ATE when X is randomized and binary; but the results on what the 'YonX estimand' (argh) is when X is continuous seem less well appreciated
macartan.github.io/assets/htmls...
comment in response to
post
Yes lots more to do; what’s solved in yitzhaki and later work is how the “ols estimand”—-the quantity targeted by lm(y~x) or the BLP—- relates to different interpretations of ate — eg avg change in Y given changes in X averaged over diff X values.
comment in response to
post
Exactly that’s the key thing: it is a weighted average but not using weights we necessarily care about. I think this is the heart of it.
The Yitzhaki piece is very interesting in this and discusses the weights that are implied by different distributions of X
www.jstor.org/stable/1392256
comment in response to
post
I don’t think I am assuming any of these things
The point here is not about the estimation method or about fitting the cef but only to highlight how wrong you can go if use reg y x to estimate an ate when x is continuous
comment in response to
post
No it just came up when I was working on something else and had me puzzling about how the OLS estimand relates to the ate in the bivariate continuous case. But clearly this is already long since solved even if not talked about enough
comment in response to
post
đź‘Ť
comment in response to
post
If estimand is the local marginal effects weighted by x density then distribution of x affects the *estimand*
If estimand is simple avg of marg effects at each x level then this equiv E[Y(max(x)-Y(min(x)] but OLS will ofc compromise to reduce error for any intermediate X vals and not just extremes
comment in response to
post
Thanks! What paper would you point to for the continuous treatment case?
comment in response to
post
Idea here is about OLS for estimation of an ate rather than fitting CEFs. But yes this is a bad fit and an alternative approach is to fit flexibly and then infer quantities of interest
comment in response to
post
Need not be but simplest to think of this as exact, one unit at each point, smooth potential outcomes function, no error
comment in response to
post
As we put it here, too often political scientists are doing parachute experiments from grounded aircraft without even knowing it
macartan.github.io/assets/pdf/p... with anna Wilke
comment in response to
post
3. Interest in whether voters evaluate candidates with feature X differently
1. Might be done with a field experiment
2. … by an information experiment embedded in a survey
3. By a survey experiment (or perhaps just a survey questin)
For 3 you might simultan’ly ask about candidates w/ and w/out X
comment in response to
post
Hi Mike it’s not about control it’s about whether the aim is to infer an effect or to measure sthg. Sometimes treatments are delivered via surveys. But v often not
Compare
1. Interest in the effect on candidates from doing X
2. Interest in the effect on voters about learnng X about a candidate…
comment in response to
post
Haha. It’s true. Sorry I deleted the previous one to fix the thread
comment in response to
post
Or: am I doing this because I really face the fundamental problem of causal inference? Or is it rather for efficiency or to address measurement issues?
The effect / measurement distinction is not about external validity but about the nature of the estimand
comment in response to
post
Hi John I think that summary is misleading
The study is not looking at the effects of actions by academics.
It looks at responses to fictitious profiles.
It’s an interesting piece for sure but subject I think to over interpretation