Anton Strezhnev
@astrezh.bsky.social
Assistant Professor, UW-Madison Political Science. http://www.antonstrezhnev.com
What the hell is going on here?
November 7, 2025 at 5:15 PM
What the hell is going on here?
From my intro lecture to the undergrads
September 7, 2025 at 3:04 PM
From my intro lecture to the undergrads
Loving my new Gmail promotions tab
August 23, 2025 at 2:28 AM
Loving my new Gmail promotions tab
Yeah, they’re winging it
July 31, 2025 at 2:10 PM
Yeah, they’re winging it
Lando Calrissian references are frankly too generous b/c of Billy Dee cool - I prefer Simon Callow’s bit part from the JCVD Street Fighter movie.
July 29, 2025 at 2:35 PM
Lando Calrissian references are frankly too generous b/c of Billy Dee cool - I prefer Simon Callow’s bit part from the JCVD Street Fighter movie.
Love this awesome/depressing paper by zikai.li - more evidence of the unfortunate failures of “deliverism” - places that were *just* eligible for a massive IRA green energy investment program were actually more likely to vote for Trump in 2024! tiny.cc/energycommunities
July 18, 2025 at 11:16 PM
Love this awesome/depressing paper by zikai.li - more evidence of the unfortunate failures of “deliverism” - places that were *just* eligible for a massive IRA green energy investment program were actually more likely to vote for Trump in 2024! tiny.cc/energycommunities
And in fact, we find no distinguishable pre-trends in the western subsample compared to the east...but we also find no evidence for treatment effects. (12/14)
July 11, 2025 at 6:12 PM
And in fact, we find no distinguishable pre-trends in the western subsample compared to the east...but we also find no evidence for treatment effects. (12/14)
There's a lot in the appendix looking at adjustments for linear pre-trends, why matching on pre-trends is likely problematic (regression-to-the-mean bias), synthetic controls and more. (10/14)
July 11, 2025 at 6:12 PM
There's a lot in the appendix looking at adjustments for linear pre-trends, why matching on pre-trends is likely problematic (regression-to-the-mean bias), synthetic controls and more. (10/14)
But one other problem with both of these approaches is that the placebos are hard to interpret - they're averaging over both short and long DiDs. Typically event study plots fix one baseline period for all estimates. When we do this, the linear pre-trend from about 1988/1992 is hard to miss. (9/14)
July 11, 2025 at 6:12 PM
But one other problem with both of these approaches is that the placebos are hard to interpret - they're averaging over both short and long DiDs. Typically event study plots fix one baseline period for all estimates. When we do this, the linear pre-trend from about 1988/1992 is hard to miss. (9/14)
We re-do the analysis using the leave-one-out estimator and show that point estimates + confidence intervals for the pre-periods roughly double. By the same equivalence testing standards, we'd fail to conclude that the pre-trend estimates are "small" (8/14)
July 11, 2025 at 6:12 PM
We re-do the analysis using the leave-one-out estimator and show that point estimates + confidence intervals for the pre-periods roughly double. By the same equivalence testing standards, we'd fail to conclude that the pre-trend estimates are "small" (8/14)
The original analysis in the paper used fect + in-sample imputation and concluded that although there was some evidence for pre-trends (CIs don't include zero), the actual magnitudes were negligible compared to the treatment effects (using an equivalence testing approach). (7/14)
July 11, 2025 at 6:12 PM
The original analysis in the paper used fect + in-sample imputation and concluded that although there was some evidence for pre-trends (CIs don't include zero), the actual magnitudes were negligible compared to the treatment effects (using an equivalence testing approach). (7/14)
When is this a problem? We show that the key factors that drive the attenuation bias are having a small share of controls to treated units and a small number of pre-treatment periods. Interestingly *more* staggering in treatment makes this problem *less* acute. (5/14)
July 11, 2025 at 6:12 PM
When is this a problem? We show that the key factors that drive the attenuation bias are having a small share of controls to treated units and a small number of pre-treatment periods. Interestingly *more* staggering in treatment makes this problem *less* acute. (5/14)
New working paper! - zikai.li and I look at the problem of assessing pre-trends using one of the "new" DiD methods - fixed effects regression imputation. tl;dr - don't use the same regression you used to impute post-treatment to also impute pre-treatment. osf.io/preprints/so... (1/14)
July 11, 2025 at 6:12 PM
New working paper! - zikai.li and I look at the problem of assessing pre-trends using one of the "new" DiD methods - fixed effects regression imputation. tl;dr - don't use the same regression you used to impute post-treatment to also impute pre-treatment. osf.io/preprints/so... (1/14)
July 10, 2025 at 3:13 PM
It also matches my preferred way of describing DiD as a difference in two cross-sections
June 9, 2025 at 2:04 PM
It also matches my preferred way of describing DiD as a difference in two cross-sections
It’s about a 2x distance, so the Cybertrucks likely get abandoned in Nashville
June 5, 2025 at 8:13 PM
It’s about a 2x distance, so the Cybertrucks likely get abandoned in Nashville
