Should we use the firm and industry fixed effects together?

[Phuc Nguyen]

I am wondering if we can use the firm and industry fixed effects together in a regression (panel data).

In particular, in a regression, whether we should use such a code in STATA

 
 reghdfe y x, a( firm industry)

where firm is the variable of firms and industry is the variable standing for industries.

I think we can use both fixed effects because it is like we control for the variables at firm and industry levels that have not yet been in the equation. However, oppositely, I wonder if the industry fixed effect is redundant here because it may be overlapped by the firm fixed effects?

[Andrew Musau]

If firms do not change industries, the industry effects are collinear with the firm effects and you can drop them.

[Phuc Nguyen]

Thank you Andrew Musau

[Maxence Morlet]

I am wondering if we can use the firm and industry fixed effects together in a regression (panel data).

In particular, in a regression, whether we should use such a code in STATA

reghdfe y x, a( firm industry)

where firm is the variable of firms and industry is the variable standing for industries.

I think we can use both fixed effects because it is like we control for the variables at firm and industry levels that have not yet been in the equation. However, oppositely, I wonder if the industry fixed effect is redundant here because it may be overlapped by the firm fixed effects?

Dear Phuc Nguyen, I would add time fixed-effects as well if I were you.

Concerning standard errors and the level of clustering, how many firms and how many industries di you have?

[Phuc Nguyen]

Dear Phuc Nguyen, I would add time fixed-effects as well if I were you.

Concerning standard errors and the level of clustering, how many firms and how many industries di you have?

Hi Maxence Morlet , I did not control for time fixed effects due to that I have a global index in my dataset (fully autocorrelated with the year fixed effects)
I have around 40k firms and 20 industries in my sample.

[Maxence Morlet]

So your global index fully captures the time variation in your data?

If you have only 20 industries, I would cluster standard errors by firm. With 40 000 firms, in relatively balanced panel data, standard errors should be consistent as 40 000 is close enough to infinity to trigger the asymptotic properties of cluster-robust standard errors.