Is there a conflict between the industry effect and firm fixed effect？

[Johnny Cash]

Hi guys, the model i built is xtreg y x1 x2 x3 i.year i.industry_code,fe
I'm trying to control the year effect, industry effect and firm fixed effect. But a reviewer pointed out that if the industry effect and firm fixed effect are included in a regression simultaneously, the industry effect will be omitted, since the industry effect and firm fixed effect are both time-invariant. It sounds right because the estimation is within-group.
However, the question is the results are different when I run

xtreg y x1 x2 x3 i.year i.industry_code,fe

and

xtreg y x1 x2 x3 i.year ,fe

like the coefficients of independent variables from the two regressions are different

Then how to explain the difference if the industry effect will be omitted in the estimation?

[Maarten Buis]

If a company changes industry, then the industry might have an effect. So that would be the first thing I would check. Such a change could be a real change in the company. It could also be that a company could be categorized in multiple industries, in in one year a coder chose one industry and in another anaother coder chose another. The change in industrycode could also be a simple typo. So I would find the companies in your data that changed industry, and see which scenario applies. After that you can make an informed decision.

[Johnny Cash]

If a company changes industry, then the industry might have an effect. So that would be the first thing I would check. Such a change could be a real change in the company. It could also be that a company could be categorized in multiple industries, in in one year a coder chose one industry and in another anaother coder chose another. The change in industrycode could also be a simple typo. So I would find the companies in your data that changed industry, and see which scenario applies. After that you can make an informed decision.

Thanks for your advice. The situation u mentioned doesn't exist in my dataset, which means the company doesn't change industry in the sample period. So I am still confused about the problem.

[Maarten Buis]

If you have fixed effects on companies and companies don't change industry, then the industry indicator variables have to be kicked out of your model. Is that the case? I am asking because you only said that the effects of the other variables changed, not whether or not the model gave effects for the industry indicator variables.

[Phil Bromiley]

To expand on Maarten's comment, the fixed effects for firm control for all omitted panel-level variables that do not vary within firm. Since industry doesn't vary within firm, firm fully controls for industry. This is why you can't include industry since it is collinear or fully controlled for by the fixed effect on firm.

[Johnny Cash]

To expand on Maarten's comment, the fixed effects for firm control for all omitted panel-level variables that do not vary within firm. Since industry doesn't vary within firm, firm fully controls for industry. This is why you can't include industry since it is collinear or fully controlled for by the fixed effect on firm.

I agree with u that firm fixed effect fully controls for industry fixed effect. And in theory, if firm fixed effect is in the model, the industry fixed effect should not be included, because it's unnecessary except the situation Maarten mentioned.
What I'm confused is that the results from

xtreg y x1 x2 x3 i.year i.industry_code,fe

and

xtreg y x1 x2 x3 i.year ,fe

are different, like coefficents and standard erroes of x1 x2 x3, and i.industry_code is not omitted in the result.

[Maarten Buis]

So you have a problem with your data: either data is setup to not do firm fixed effects (review the output from xtset) or your industry_code does change for at least one firm (review the output from xtsum industry_code, you want the within standard deviation to be exactly 0).