Right now’s put up will share some footage from Berlin, but additionally decide up the place an earlier put up on covariates in diff-in-diff left off. This time, I simply needed to mix formalism with some ideas I’ve. I’ll present you the derivation for when covariate imbalance will break parallel traits and when it received’t. After which I’ll simply share my opinions on critiquing the weak spot of a diff in diff. So in case you are enthusiastic about both test it out.
However I additionally flipped a coin 3x (or had python do it anyway), and it got here up two heads out of three which implies this’ll be paywalled. Thanks once more to your assist!
Guten tag! After 4 days and three nights in lovely Berlin. I’m heading to the airport the place I’ll hop on a aircraft to Pisa — again to Tuscany! I simply left there! — after which e-book it over to Lucca. I spent two weeks in Lucca, Italy final 12 months and had such a pleasant time. I’ll be presenting my paper on discretion and the conduct of AI brokers. Ask IMT about it and see in the event that they’ll allow you to peek your head in!
However I had a fully unbelievable time on the Berlin Faculty of Economics. I offered at DIW Berlin, and met new individuals. It was actually superb. We’re speaking about doing it once more subsequent 12 months even.
However for now, I transfer on. I’ve two extra issues earlier than I’m carried out and head to San Sebastián within the Basque Nation, recognized for its fabulous concha bay, superb meals and artificial management. This weeks it’s IMT in Lucca, after which beginning this weekend, Belgium, to KU Leuven.
However let me pause there, as a result of what I wish to do now could be decide up on a substack from final week and present you formally the conditions the place it’s obligatory to incorporate coverages in a difference-in-differences.
The core purpose is that you simply want the covariates to fulfill conditional parallel traits, which we talk about in part 4 of our JEL. There’s three separate bias phrases although. And they’re:
-
The covariates it is advisable fulfill conditional parallel traits are imbalanced between therapy and management.
-
Heterogenous therapy results throughout the size of these covariates.
-
Conditional parallel traits relies on time INVARIANT covariates.
It’s fairly widespread to listen to somebody say that they don’t want to regulate for covariates as a result of the whole lot that determines the end result that doesn’t change over time will get deleted with the primary distinction. Or, put one other method, it’s getting absorbed within the unit fastened impact.
However this isn’t correct. Diff-in-diff is powerful to variables that decide Y(0), the extent of the end result, however recall the bias of diff-in-diff: it isn’t based mostly on degree variations. It’s based mostly on pattern variations in Y(0).
Thus in case you want them to fulfill conditional parallel traits, then after all deleting them received’t assist. You’ll want to determine a technique to hold them in.
Properly associated to that point invariant case is the extra normal drawback: imbalance. And this time I wish to present you formally what occurs, as final time it was solely a easy numerical instance. However I get pleasure from working from the regression equation to the formal depiction of the issue and I wager others do as nicely.
Let’s begin out with a two individuals mannequin. Since we’re together with covariates, we can’t merely write it down as 4 averages and three subtractions, so let’s as an alternative write it down as a two method fastened results regression with additive covariates as an instance the issue with “altering returns to the covariate over time”. Right here is that regression within the first interval, interval 1.
(Y(0)_{1,i} = alpha_i + tau_1 + beta_1 X_{1,i} + varepsilon_{1,i} )
And right here it’s in interval 2.
(Y(0)_{2,i} = alpha_i + tau_t + beta_2 X_{2,i} + varepsilon_{2,i})
So discover that I put subscripts 1 and a couple of on beta. What does that imply? That implies that the impact in interval 1, beta_1, could or will not be what it grew to become later in interval 2, beta_2.
