Sven Resnjanskij

50 posts

Sven Resnjanskij banner
Sven Resnjanskij

Sven Resnjanskij

@SvenRes

PostDoc, Research Network Affiliate at CESifo Munich

Stockholm, Sweden Katılım Mart 2012
176 Takip Edilen108 Takipçiler
Sven Resnjanskij retweetledi
Ludger Woessmann
Ludger Woessmann@Woessmann·
🚀Mentoring durch ehrenamtliche Studierende verdoppelt ‼️ die Chance von benachteiligten Jugendlichen, eine Berufsausbildung zu beginnen👍 Unsere Evaluierung der Effekte von @RYL_Mentoring Ergebnisse jetzt im @ifo_Institut Schnelldienst: 👉ifo.de/publikationen/… 🧵 1/6
GIF
Deutsch
5
16
46
5.2K
Sven Resnjanskij
Sven Resnjanskij@SvenRes·
@hooster1 @KLM Hope the Deutsche Bahn night-train substitute worked ok! That can be quite an experience for itself :)
English
0
0
0
0
Hessel Oosterbeek
Hessel Oosterbeek@hooster1·
Retourvlucht Amsterdam-München met @KLM. Heenvlucht op donderdag vertrekt met 2 uur vertraging. Terugvlucht op zondag geannuleerd. Vervangende vlucht pas op maandagochtend. Gaat lekker na twee jaar aan het infuus van de staat.
Nederlands
2
0
3
0
Sven Resnjanskij retweetledi
Susan Athey
Susan Athey@Susan_Athey·
After 5 yrs of iterating on math, simulations, writing, we updated our working paper on clustering. Lots of time working on single formulation that captures intermediate cases, simple after-the-fact, math matches intuition. arxiv.org/abs/1710.02926
English
2
58
253
0
Sven Resnjanskij retweetledi
Johannes Haushofer
Johannes Haushofer@jhaushofer·
I've been in touch with several German universities which have expressed interest in admitting students displaced from Ukraine. If you're a displaced student, you can register your interest via this form, and I will share it with the universities: forms.gle/S2f4QURoVJWaM1…
English
13
260
361
0
Sven Resnjanskij retweetledi
Jeffrey Wooldridge
Jeffrey Wooldridge@jmwooldridge·
@SvenRes @lewbel But with moderate trends they can differ, with logit or probit being better across different DGPs. With strong trends (not violation of PT, BTW), the LPM can even give the wrong sigh. Logit, probit never do -- so far.
English
0
2
2
0
Sven Resnjanskij
Sven Resnjanskij@SvenRes·
@jmwooldridge @lewbel Am I wrong, assuming that in the binary-response DiD case, it would be an option to still run a linear DiD spec., when it'S feasable to express all RHS variables in terms of binary variables as well? My intuition is, that (except SE) OLS would give similar estimates.
English
1
1
0
0
Jeffrey Wooldridge
Jeffrey Wooldridge@jmwooldridge·
@lewbel It is a very strong assumption. And it's especially strong when the linear form is applied to, say, a binary response. So those who say "just use OLS" seem not to realize that a linear model can be very poor in these situations. So I want to try it a few different ways.
English
2
1
7
0
Jeffrey Wooldridge
Jeffrey Wooldridge@jmwooldridge·
Here's a simulation with T = 2, no X. PT holds in logs. Mean SD ATT .7703254 .2274547 Linear .0732538 .2850102 Expon .7704271 .2405215 The linear is badly biased. I'll post the simulation on my Dropbox.
Jeffrey Wooldridge@jmwooldridge

@jondr44 @instrumenthull @EpiEllie In the exponential case, PT is applied to log[E(Y(0)|D,X)]. Is this a better assumption than levels? I think so. But it's hardly perfect. By using linear and exponential we can hope to bound the effect. But we might not like the bounds.

English
3
5
52
0
Sven Resnjanskij
Sven Resnjanskij@SvenRes·
@lewbel When I, as a young PhD student, cited your paper to make the case for a Probit instead of LPM in front of an “Mostly-Harmless”-labor-econ audience, no one believed in the sign change. So I showed some Stata output. I still got the comment to estimate a LPM...
Sven Resnjanskij tweet media
English
0
0
1
0
Arthur Lewbel
Arthur Lewbel@lewbel·
In this paper we showed that the linear probability model could estimate a negative ATE even if every person in the population actually had a nonnegative treatment effect. Maybe should construct similar for linear Diff-in-Diff. doi.org/10.1111/j.1540…
English
1
1
21
0
Arthur Lewbel
Arthur Lewbel@lewbel·
Then: arguing for or against the linear probability model. Now: arguing for or against parallel trends with a binary outcome. #EconTwitter
Jeffrey Wooldridge@jmwooldridge

@lewbel It is a very strong assumption. And it's especially strong when the linear form is applied to, say, a binary response. So those who say "just use OLS" seem not to realize that a linear model can be very poor in these situations. So I want to try it a few different ways.

English
1
3
21
0
Sven Resnjanskij retweetledi
Arthur Lewbel
Arthur Lewbel@lewbel·
Hot take: 2022 will be the year of backlash against Diff-in-Diff models. Users will finally realize parallel trends is often an implausible, structural assumption. E.g., What stops control group agents from modifying behavior based on observing the treated group? #EconTwitter
English
7
16
157
0
Sven Resnjanskij retweetledi
Jeffrey Wooldridge
Jeffrey Wooldridge@jmwooldridge·
As another analogy, we now know from @TymonSloczynski's nice REStat paper on regression adjustment that OLS of Y on 1, D, X can produce badly biased estimators of ATE and ATT. But full RA, Y on 1, D, X, D*(X - Xbar) can work very well (with overlap).
Jeffrey Wooldridge@jmwooldridge

As I've emphasized before on here and in my TWFE/TW Mundlak paper, we shouldn't be blaming an estimation method for our lack of imagination in modeling. For binary treatment, staggered intervention, TWFE will full heterogeneity in cohort, time, covariates does what we want.

English
3
11
67
0

Keşfet

@unierfurt @JRuhose @kieluni @katharina_wedel @LMU_Muenchen @SimonWiederhold @IWH_Halle @Woessmann