[Stanislav Beli]

Thank you Jeff for a quick reply.

Treatment_it stands for variable that equals 1 if bank i at time t has the factor variable (factor1) greater than or equal to the chosen threshold and takes value 0 if it is lower than threshold - therefore there is a subscript t as well. Is it wrong though ?

Basically to say it more simple : In case bank i at the specific time t = 2013 has factor variable (factor1) greater than or equal to the chosen threshold then that bank i will take value 1 for the time t (2012,2013,2014,2015,2016,2017,2018 and 2019) and takes value 0 for the time t (2012,2013,2014,2015,2016,2017,2018 and 2019) if it is lower than threshold

This is how I made 2 different samples according to the condition from 2013.

Also I have used it as you said - two-way FE with just the interaction and I got the same results as with the use of the equation I wrote above including Treatment and PostPeriod as well as interaction.

What you think regarding those omitted observations - year 2019 and 99th id ? = I have tried this: reg NIM Treat_PostPeriod Banksize Depositratio RWA Liquidity Creditrisk GDP INF HHI i.Year i.id, cluster (id) and there are no omitted observations - do you think it was due to that ?

Thank you very much.

[Jeff Wooldridge]

Stanislav: The equation you wrote down and your Stata code are inconsistent. The variable Treat that you defined does not vary with time. It can't, because then it would identical to Treat_PostPeriod, so one of them would drop out. You defined Treat in the usual way: it is one if the unit is eventually treated and zero if not. Therefore, it does not change over time.

I'm certain that the reason there are no omitted variables (not observations, variables) with

reg NIM Treat_PostPeriod Banksize Depositratio RWA Liquidity Creditrisk GDP INF HHI i.Year i.id, cluster (id)

is because you've dropped the two redundant variables Treat (doesn't change across t) and PostPeriod (doesn't change across i).

[Stanislav Beli]

Thank you very much for a quick response.

I have corrected the equation as you said so the Stata code would be reflecting what is written.

Although after the regression I have a problem which I can not really solve : when I run the regression
reg NIM Treat_PostPeriod Banksize Depositratio RWA Liquidity Creditrisk GDP INF HHI not including fixed effects neither for banks nor for the years and not even clustering the errors I get that almost every variable is significant (2/8 are insignificant).

Although when I run it as I stated before with i.Year i.id, cluster (id) I get half variables insignificant.

Can you maybe explain it why is it like that and whether I should use RE instead of FE or something else ?

Thank you very much.

[Carlo Lazzaro]

Stanislav:
I find difficul to believe that default standard errors are the way to go with your regression.
That said, you can easily see with -linktest- whether the functional form of the regressand is well specified or not.
A more general comment: statistical significance is simply a feature of your results. It's much more impoirtant to give a fair and true view of th data generating process you're interested in instead of hunting for p<0.05 (see also

https://www.amstat.org/asa/files/pdfs/p-valuestatement.pdf

).

[Jeff Wooldridge]

Stanislav: When you don't use fixed effects you must include Treat and PostPeriod. It's only when you include unit and time fixed effects that those become redundant. If your covariates didn't change over time the estimates would be identical. With time-varying covariates they are different. I have a recent paper that discusses all of this, and I can send you a link if you'd like. I would stick with the fixed effects estimation but hopefully using OLS and putting in Treat and PostPeriod gives similar results. You should continue to cluster your standard errors.