Using i.year and pre/post-dummy and in FE panel regression

[Eivind Andersen]

Hi,

I'm estimating a fixed effects panel regression to study whether the implementation of IFRS 9 in 2018 affected earnings management in EU banks. I have a policy dummy variable, IFRS9, which equals 1 for the years 2018–2023 and 0 for 2012–2017.

My main concern is whether it's appropriate to include both i.year and the IFRS9 dummy in the same FE model. Since IFRS9 varies only across time and not across banks, it appears to be perfectly collinear with the year fixed effects. When I include i.year, Stata drops the 2023 year dummy but not IFRS9 "year.2023 omitted because of collinearity". Does this mean the effect of IFRS9 is being absorbed by the year dummies, and that the coefficient on IFRS9 is not really identified?

I'm also including a macroeconomic variable, GDP per capita, which varies by country and year, but is constant across banks within each country-year. Could this also become collinear with i.year, especially since GDP is likely correlated with time-specific factors? What's the best practice in this kind of setup?



Stata input and output:
xtreg LLPL_w L_NPLL_w CHANGENPLL_w CHANGELOANSL_w EBTLLPL_w NEWTOTALCAP BNP EBTLLPL_w_IFRS9 i.IFRS9 i.year, fe vce(cluster firm_id)



Thanks in advance,
Eivind

[Eivind Andersen]

I can't seem to copy the output without messing up the table

[Eivind Andersen]

here it is

[Carlo Lazzaro]

Eivind:
why not considering -xtdidregress-?

[Eivind Andersen]

Thanks for the reply Carlo!

I must confess I am not familiar with the function, could you please explain? I also don't have a control group, so Im not sure if difference in differences is appropriate? bear in mind, Im a masters student, with limited Stata-experience, sorry for this.

Best regards, Eivind

[Sebastian Kripfganz]

To answer the initial question: Yes, the IFRS9 dummy is perfectly collinear with the set of time dummies and its coefficient is therefore not identified.

Regarding GDP, as long as you have variation across countries, this won't be perfectly collinear with the time dummies (which are constant across countries).

[Eivind Andersen]

Thank you so much Sebastian! I did a vif-test with and without time dummies, IFRS9 was initially 9 and fell to 1.93, BNP from 3.5 to 1,02.