Findings from the Chiu et al. study on TWFE shown here to generalize to economics.
How different are different differences-in-differences? The answer may (or may not) surprise you
New working paper, building on a dumb tweet from earlier this month x.com/instrumenthu...
A short summary follows:
New working paper, building on a dumb tweet from earlier this month x.com/instrumenthu...
A short summary follows:
October 29, 2026 at 4:29 PM
Findings from the Chiu et al. study on TWFE shown here to generalize to economics.
The last few years have been, uh, interesting for those who like to run TWFE
We now understand that w/staggered adoption, TWFE can put negative weights on heterogeneous effects - threatening sign reversals
Seems bad! But is it? The prevalence/importance of negative weights is an empirical question
We now understand that w/staggered adoption, TWFE can put negative weights on heterogeneous effects - threatening sign reversals
Seems bad! But is it? The prevalence/importance of negative weights is an empirical question
October 29, 2026 at 3:53 PM
The last few years have been, uh, interesting for those who like to run TWFE
We now understand that w/staggered adoption, TWFE can put negative weights on heterogeneous effects - threatening sign reversals
Seems bad! But is it? The prevalence/importance of negative weights is an empirical question
We now understand that w/staggered adoption, TWFE can put negative weights on heterogeneous effects - threatening sign reversals
Seems bad! But is it? The prevalence/importance of negative weights is an empirical question
my particular type of hell where I get rejected in field journals for not having data that can completely nail sky-high causal identification standards while top methodologists out there make the point that TWFE is actually mostly ok
October 29, 2026 at 7:14 PM
my particular type of hell where I get rejected in field journals for not having data that can completely nail sky-high causal identification standards while top methodologists out there make the point that TWFE is actually mostly ok
October 30, 2026 at 11:04 AM
Nevertheless, there are indeed sign flips in some specs and we can predict them
Earlier pubs, papers that only report TWFE, and specifications with more always-treated units / fewer never-treated units are more likely to have sign-flips
Journal prestige / paper citations are not predictive
Earlier pubs, papers that only report TWFE, and specifications with more always-treated units / fewer never-treated units are more likely to have sign-flips
Journal prestige / paper citations are not predictive
October 29, 2026 at 3:57 PM
Nevertheless, there are indeed sign flips in some specs and we can predict them
Earlier pubs, papers that only report TWFE, and specifications with more always-treated units / fewer never-treated units are more likely to have sign-flips
Journal prestige / paper citations are not predictive
Earlier pubs, papers that only report TWFE, and specifications with more always-treated units / fewer never-treated units are more likely to have sign-flips
Journal prestige / paper citations are not predictive
First, II don't think we should abandon modern DiD estimators or the lessons of the papers that derived them (including outside the negative weights concern). It's important to know what TWFE estimates, what the potential problems are, and how to check for them in any given application
That said...
That said...
October 29, 2026 at 3:59 PM
First, II don't think we should abandon modern DiD estimators or the lessons of the papers that derived them (including outside the negative weights concern). It's important to know what TWFE estimates, what the potential problems are, and how to check for them in any given application
That said...
That said...
How do we separate variation due to estimation error vs. estimands? I estimate a hierarchical model:
- TWFE's estimand is 91-98% correlated w/modern DiD estimands
- The probability of a sign reversal is 6-12%
- Probability of a *significant* wrong-signed TWFE is 0.6-2.1%
What's the takeaway?
- TWFE's estimand is 91-98% correlated w/modern DiD estimands
- The probability of a sign reversal is 6-12%
- Probability of a *significant* wrong-signed TWFE is 0.6-2.1%
What's the takeaway?
October 29, 2026 at 3:58 PM
How do we separate variation due to estimation error vs. estimands? I estimate a hierarchical model:
- TWFE's estimand is 91-98% correlated w/modern DiD estimands
- The probability of a sign reversal is 6-12%
- Probability of a *significant* wrong-signed TWFE is 0.6-2.1%
What's the takeaway?
- TWFE's estimand is 91-98% correlated w/modern DiD estimands
- The probability of a sign reversal is 6-12%
- Probability of a *significant* wrong-signed TWFE is 0.6-2.1%
What's the takeaway?
I use the latest GPT models to search + replicate staggered DiD specs from ten general-interest journals over the last 20+ years and compare TWFE vs modern DiD estimators
I get a sample of 1,300+ specs from 100+ papers
How correlated are the estimators & how many sign flips do we end up seeing?
I get a sample of 1,300+ specs from 100+ papers
How correlated are the estimators & how many sign flips do we end up seeing?
October 29, 2026 at 3:55 PM
I use the latest GPT models to search + replicate staggered DiD specs from ten general-interest journals over the last 20+ years and compare TWFE vs modern DiD estimators
I get a sample of 1,300+ specs from 100+ papers
How correlated are the estimators & how many sign flips do we end up seeing?
I get a sample of 1,300+ specs from 100+ papers
How correlated are the estimators & how many sign flips do we end up seeing?
also, panel data simply means measuring the same units repeatedly, and can be expressed in either wide or long format. But many people mistakenly think of panel data only as the long format, often equating Difference-in-Differences solely with long format-based TWFE regressions.
October 21, 2026 at 4:26 PM
also, panel data simply means measuring the same units repeatedly, and can be expressed in either wide or long format. But many people mistakenly think of panel data only as the long format, often equating Difference-in-Differences solely with long format-based TWFE regressions.
Have we abandoned #TWFE too quickly?😮
New in #Sociological #Methodology w/ @aksoyundan.bsky.social
We compare TWFE event studies vs dynamic #DID estimators. Results suggest a more nuanced picture than "never use TWFE" and clarify when different estimators perform well 👇
🔗 doi.org/10.1177/0081...
New in #Sociological #Methodology w/ @aksoyundan.bsky.social
We compare TWFE event studies vs dynamic #DID estimators. Results suggest a more nuanced picture than "never use TWFE" and clarify when different estimators perform well 👇
🔗 doi.org/10.1177/0081...
October 15, 2026 at 2:32 PM
Have we abandoned #TWFE too quickly?😮
New in #Sociological #Methodology w/ @aksoyundan.bsky.social
We compare TWFE event studies vs dynamic #DID estimators. Results suggest a more nuanced picture than "never use TWFE" and clarify when different estimators perform well 👇
🔗 doi.org/10.1177/0081...
New in #Sociological #Methodology w/ @aksoyundan.bsky.social
We compare TWFE event studies vs dynamic #DID estimators. Results suggest a more nuanced picture than "never use TWFE" and clarify when different estimators perform well 👇
🔗 doi.org/10.1177/0081...
Given the way that the DiD literature, it was kind of inevitable that TWFE would come back from the dead.
More to come... 👀
More to come... 👀
Have we abandoned #TWFE too quickly?😮
New in #Sociological #Methodology w/ @aksoyundan.bsky.social
We compare TWFE event studies vs dynamic #DID estimators. Results suggest a more nuanced picture than "never use TWFE" and clarify when different estimators perform well 👇
🔗 doi.org/10.1177/0081...
New in #Sociological #Methodology w/ @aksoyundan.bsky.social
We compare TWFE event studies vs dynamic #DID estimators. Results suggest a more nuanced picture than "never use TWFE" and clarify when different estimators perform well 👇
🔗 doi.org/10.1177/0081...
October 15, 2026 at 4:15 PM
Given the way that the DiD literature, it was kind of inevitable that TWFE would come back from the dead.
More to come... 👀
More to come... 👀
Did you see P Hull's recent exercise?
github.com/peterdhull/t...
github.com/peterdhull/t...
GitHub - peterdhull/twfe-replication
Contribute to peterdhull/twfe-replication development by creating an account on GitHub.
github.com
October 15, 2026 at 6:58 PM
Did you see P Hull's recent exercise?
github.com/peterdhull/t...
github.com/peterdhull/t...
Ahhh i should have been saying TWFE probably, as I think that's ultimately what would be more applicable.
October 12, 2026 at 11:10 PM
Ahhh i should have been saying TWFE probably, as I think that's ultimately what would be more applicable.
It's specified as something like lm(y ~ post + treat + post:treat), but it is worth learning for the assumptions behind it and for its generalizations/extensions (eg, TWFE, synthetic control).
October 12, 2026 at 8:24 PM
It's specified as something like lm(y ~ post + treat + post:treat), but it is worth learning for the assumptions behind it and for its generalizations/extensions (eg, TWFE, synthetic control).
If anything - all of the new methods are just versions of TWFE anyway! Just specify the TWFE correctly and any concerns about TWFE disappear :)
October 10, 2026 at 3:21 PM
If anything - all of the new methods are just versions of TWFE anyway! Just specify the TWFE correctly and any concerns about TWFE disappear :)
More generally, I think that all this back and forth about "TWFE" feeds into the general misunderstanding that "causal inference" is somehow different than "regression with a clear estimand and useful control variables."
October 11, 2026 at 1:43 PM
More generally, I think that all this back and forth about "TWFE" feeds into the general misunderstanding that "causal inference" is somehow different than "regression with a clear estimand and useful control variables."
That’s useful! But it’s very different from telling Us how much actual people’s results differ across TWFE/staggered methods in practice.
October 11, 2026 at 1:00 PM
That’s useful! But it’s very different from telling Us how much actual people’s results differ across TWFE/staggered methods in practice.
I happen to believe that our simulation paper does something very similar: it shows how much effect heterogeneity affects the results in TWFE and the other DID estimators + compares this bias to biases from other violations across all estimators 🤷♂️
October 11, 2026 at 8:05 AM
I happen to believe that our simulation paper does something very similar: it shows how much effect heterogeneity affects the results in TWFE and the other DID estimators + compares this bias to biases from other violations across all estimators 🤷♂️
Less “BLUE” and more “the F-test on the restrictions is usually not significant”
All the new methods are also just TWFE with more parameters (and occasionally IPW weights)
All the new methods are also just TWFE with more parameters (and occasionally IPW weights)
October 10, 2026 at 3:24 PM
Less “BLUE” and more “the F-test on the restrictions is usually not significant”
All the new methods are also just TWFE with more parameters (and occasionally IPW weights)
All the new methods are also just TWFE with more parameters (and occasionally IPW weights)
The issue here is that it’s not “apples-to-apples” - saying that you can fix static TWFE bias due to time heterogeneity by fitting the dynamic is already acknowledging that we needed a different estimator.
October 10, 2026 at 3:20 PM
The issue here is that it’s not “apples-to-apples” - saying that you can fix static TWFE bias due to time heterogeneity by fitting the dynamic is already acknowledging that we needed a different estimator.
Marianne Bitler suggested to me recently that in practice TWFE and the new staggered treatment event history approaches produce very similar estimates in general. So this tracks.
Going to use this to make the requisite BLUE jokes though.
Going to use this to make the requisite BLUE jokes though.
October 10, 2026 at 3:03 PM
Marianne Bitler suggested to me recently that in practice TWFE and the new staggered treatment event history approaches produce very similar estimates in general. So this tracks.
Going to use this to make the requisite BLUE jokes though.
Going to use this to make the requisite BLUE jokes though.
In reality it is good to make the scope conditions explicit like this. Even if most of the time standard TWFE is basically fine
October 10, 2026 at 3:59 PM
In reality it is good to make the scope conditions explicit like this. Even if most of the time standard TWFE is basically fine
I skimmed this and it is not nearly as good as the recent paper in SM. I recommend Ruttenauer and Aksoy instead.
journals.sagepub.com/doi/full/10....
journals.sagepub.com/doi/full/10....
When Can We Use Two-Way Fixed Effects (TWFE)? A Comparison of TWFE and Novel Dynamic Difference-in-Differences Estimators - Tobias Rüttenauer, Ozan Aksoy, 2026
The conventional two-way fixed effects (TWFE) estimator has come under increasing scrutiny in settings with staggered treatment adoption. Recent research shows ...
journals.sagepub.com
October 9, 2026 at 3:51 PM
I skimmed this and it is not nearly as good as the recent paper in SM. I recommend Ruttenauer and Aksoy instead.
journals.sagepub.com/doi/full/10....
journals.sagepub.com/doi/full/10....
peter hull comparing TWFE and modern DIDs & it seems TWFE is OK.
people will go "YOU SEE IT DIDN'T MATTER AFTER ALL" but regardless of whether TWFE is ultimately useful or not that all DiD chaos was a net good because it clarified the estimand.
tinyurl.com/twfevsdid
people will go "YOU SEE IT DIDN'T MATTER AFTER ALL" but regardless of whether TWFE is ultimately useful or not that all DiD chaos was a net good because it clarified the estimand.
tinyurl.com/twfevsdid
October 8, 2026 at 11:31 PM
peter hull comparing TWFE and modern DIDs & it seems TWFE is OK.
people will go "YOU SEE IT DIDN'T MATTER AFTER ALL" but regardless of whether TWFE is ultimately useful or not that all DiD chaos was a net good because it clarified the estimand.
tinyurl.com/twfevsdid
people will go "YOU SEE IT DIDN'T MATTER AFTER ALL" but regardless of whether TWFE is ultimately useful or not that all DiD chaos was a net good because it clarified the estimand.
tinyurl.com/twfevsdid