xtreg, fe with or without year dummies to control for the financial crisis

[Victoria Rogers]

I'm using xtreg, fe

Should I include or exclude time dummies to control for the financial crisis?
You should never use industry and firm dummies in case of a fixed model like xtreg, fe because the effects of industry and firm dummies are then already controlled for by the fixed effects model.
However, when I searched on the Internet, I saw that some people still use year dummies while running a fixed model like xtreg, fe

[Clyde Schechter]

You should never use industry and firm dummies in case of a fixed model like xtreg, fe because the effects of industry and firm dummies are then already controlled for by the fixed effects model.

That makes no sense. If you have already -xtset- your data with industry or firm as the panel variable, then, no, you can't add dummies for industry or firm. But you can add anything you want to your model, provided it makes sense. The only thing you have to avoid is whatever panel variable you specified when you -xtset- your data, and you need to look at your other variables to see if any of them might be collinear with one or more of the time dummies. (Actually, even if you violate that rule, the only thing that will happen is that Stata will exclude your dummies from the regression because of collinearity--you will still get good results.)

Whether dummies for individual years are the best way to account for effects of the financial crisis, or some other specification (perhaps a single dummy covering the era from the start to the end of the financial crisis, or some other more complex specification that represents the "intensity" of the financial crisis in each year) is not a statistical question really: it depends on what the goals of your analysis are, what theory has to say about the nature and magnitude of the effects of the financial crisis on your outcomes, and should also take into account conventions and traditions in your field. But if you have concluded based on those considerations that individual year dummies are the best way to represent this, there is no reason you can't include them in a fixed-effects regression.

[Victoria Rogers]

Thank you Clyde. I meant: "You should never use industry and firm dummies in case of a fixed model like xtreg, fe because the fixed effects of industry and firm dummies are then already controlled for by the fixed effects model." However, it's indeed logical that random effects of industry (not the panelvar) can be controlled for with i.industry

[Carlo Lazzaro]

Victoria:
as an aside to Clyde's sound advices, the best way to make an educated guess about your query is to look at what you typed and what Stata gave you back.
Besides, you query implies that results of Hausman specification test favoured fixed vs random effect linear panel data regression.

Kind regards,
Carlo

[Carlo Lazzaro]

Victoria:
then what is the meaning of your statement

...that random effects of industry (not the panelvar) can be controlled for with i.industry

if your specifications favours fixed vs random effects?

Kind regards,
Carlo

[Victoria Rogers]

Carlo, it's logical otherwise some researchers wouldn't add i.industry in a -xtreg, fe- model, why else would they add it. So, my question still remains why would anyone add i.industry in a fixed model?

[Carlo Lazzaro]

Victoria:
I was unclear in my previous reply.
I meant that if industry is a time-invariant predictor, -xtreg, fe- (unlike -xtreg, re-) will not produce any estimate of its effect on the dependent variable.
Kind regards,
Carlo

[Victoria Rogers]

Ok thank you Carlo. If the industry rarely changes, is it in that case still a nice idea to add i.industry in a -xtreg, fe-?

[Carlo Lazzaro]

Victoria:
I do not think that adding i.industry is a nice idea if it is a time invariant predictor.

Kind regards,
Carlo

[Phil Bromiley]

This is a little simpler than it sounds. Fixed effects is essentially dummies for all the values of the panel variable. So, fixed effects for firms essentially includes dummies for each firm. Since firms very seldom change industries and firms are nested within industries, the firm fixed effect will fully pick up any stable industry effect. In other words, if you put in an industry dummy, it would be colinear with the sum of the firm fixed effects.

You can add industry with random effects because the random effects are forced to be uncorrelated with any of the variables included in the model.

You might also consider creating an annual industry average of whatever is the dv, excluding the value for the firm of interest. This should then give you a one-parameter way to control for annual effects that I would argue is better than annual dummies since it allows the influence to vary across industries.