using the i.year for time-fixed effects in a Differences-in-Differences Approach

[Johann Holst]

Hello everyone,

I am not sure if a combination of a time-fixed-efffects model and a Differences-in-Differences model is possible and if it is, how to correctly compute that within stata. Further, I wanted to confirm that my Standard Errors are clustered on Entity Level and if a wrong technical command within stata leads to my very low R².

I am researching on the impact of the introduction of the Financial Transaction Tax in France on stockmarket quality Parameters like tradingvolume and volatility. To answer this Question I choose the differences-in-differences Approach on my paneldata, using the command:

xtreg tradingvolume controls TimeDummy TreatmentDummy InteractionTermTimeTreatment , fe robust

The fixed-effects within the regression are based on my EntityID variable, so from my point of view, I control for factors that are different between enities but are constant over time.
Within the next step, I want to controll for time fixed effects, which from my point of view are factors, that are constant across different entities, but vary over time. For this I added the i.year command to my regression:

xtreg tradingvolume controls TimeDummy TreatmentDummy InteractionTermTimeTreatment i.Year , fe robust

These regressions lead to the following results:

From my perspective, the Addition of time-fixed-effects works quite well, because my SE increase without losing significance of my Interactionterm. Further my R² increased also.
Qualitativly my estimators are in line with leading literature. Another constrain regarding my Analysis for me is, that within the Literatur R² above 0,9 are reached, although they use the same Approach based on similar data. I am afraid, that I use a command technically wrong within stata and just wanted to double check.

Last Question regarding this: Using the xtreg, y x, fe robust command I get the notification "(Std. Err. adjusted for 118 clusters in EntityID)". Does this mean my Standard Errors are clustered on Entity Level?

I appreciate every Kind of help!

Johann from Hamburg

[Clyde Schechter]

You do not show the actual output of your regressions: you show them as they are laundered by -esttab- or one of the other pretty-print programs. But -xtreg, fe- produces three different R² statistics. So this raises the question of which of those you are seeing here, and which is the one used in the reports you are comparing your results to. So there isn't really anything to say about that without a great deal more information than shown in the post.

As for your other question, "

I get the notification "(Std. Err. adjusted for 118 clusters in EntityID)". Does this mean my Standard Errors are clustered on Entity Level?", yes, that is what it means, and it is telling you that there were exactly 118 such Entity Levels that contributed to the modeling.

[Johann Holst]

Hello Clyde,

thanks for catching up on this.

Here is are the results of the fixed-effects model without the timedummies, as I read in other threads in this Forum I focused on the R² within:

And here are the complete results for the fixed-effects + time-fixed-effects model:

The cited literature I meantioned was Eichfelder et al. 2017 (https://www.econstor.eu/bitstream/10.../879246561.pdf), they are using the Adjusted R², I think this is not shown in my results yet.
Another confusing Point for me is, Looking at Table II at page 48 of Eichfelder et al., is that they just write "yes" on the line "Time Fixed Effects", but do not Show any Dummies for the months.

[Clyde Schechter]

As far as I know, there is no such thing as an adjusted R² statistic for fixed-effects regression. Also, while I do not have the time to read the Eichfelder paper, and almost certainly would not understand it if I tried, doing a quick scan of it, I note that they state on numbered page 24 (PDF page 27) that they are using OLS regression. So I think that is why you are getting some rather different results.

[Carlo Lazzaro]

Johann:
as an aside to Clyde's as always enlightening advice, you might be interested in this recent thread https://www.statalist.org/forums/for...g-fixed-effect

[Johann Holst]

Clyde Schechter I was not aware that xtreg is not analog to using a OLS estimator. But I will do some further research on that topic, thanks to your advice.
Carlo Lazzaro thank you also for giving the link to the other thread. It's quite helpful.