R^2 for diff-in-diff

[Giuseppe Polito]

I'm estimating a diff-in-diff model with:

xtdidregress (y) (treated) [pweight=my_weights], group(groupnum) time(year)

Then I'm replicating the results with:

xtset groupnum year
[..]
xtreg y ib0.treated i.year [pweight=my_weights], fe vce(cluster groupnum)

I obtain the same coefficients and the same standard errors.

However, the latter model gives me an R^2, which is very low:

R-squared:                                     
     Within  = 0.0493                       
     Between = 0.0106
     Overall = 0.0026

Do I have to worry?

Do I have to look for R^2 in a diff-in-diff model at all?

[Jared Greathouse]

R2 isn't a very good metric (at least by itself!) in the best of times. Of course, there are situations where this isn't true. For DD, your concerns should be elsewhere, in my opinion.


How many treated units are there?

[Carlo Lazzaro]

Giuseppe:
as far as -xtreg,fe- is concerned, the omission of the entire Stata output makes difficult to comment on.
A very lor Rsq_within might be due to an absence of evidence of a group-wise effect (that Stata will not report with non-default standard errors).
In addition, you do not report how many panels youyr dataset is composed of: this information is relevant to understand if cluster-robust standard error is the way to go.
Eventually, did you tes whether the functional form of the regressand is correctly specified?

[Maxence Morlet]

In addition to the great advice previously given, the importance of the coefficient of determination (R squared) varies across fields. In my field for instance (microeconometrics), no one cares about the R-Squared because all the attention is on causality. If you're running a diff-in-diff, I assume you're after causal inference as well. If this is the case, then you should spend a lot more time arguing that identifying assumptions hold.

In what field do you work?