Fixed effects model

[Jan Geurst]

Hello,

I'm running a fixed effects model with several fixed effects (firm, country, industry). I have clustered the standard errors by firm.

I ran these commands:
1. reghdfe bda announcement_eligible announcement l_lnassets l_roa, absorb(c indus incorp) vce(cluster c)

2. reghdfe bda announcement_eligible announcement l_lnassets l_roa, absorb(indus incorp) vce(cluster c)

For the second regression all my coefficients become significant (where they were not in the first regression), alone the r-squared goes from 0.8969 for the first regression to 0.4429 for the second regression.

What could be a reason that I should not include firm-specific fixed effects in my regression? And why decreases my r-squared with such a large amount?

Thank you!

[Amin Sofla]

From high r-squared in the first model, and also its sharp drop in the second model, I am guessing that the firm fe is strongly correlated to that of the industry. You might want to drop industry's fe instead. If true, I would estimate it using xtreg (using i.country as a control: it does not have high dimension) because xtreg has more features.

[Jan Geurst]

Even if I drop the industry fixed effects the regressor stays the same as if you only include firm fixed effects and their t-values drop by 0.01.

Do the industry and country fixed effects have no effect in combination with firm fixed effects, because the firm fixed effects create so many groups that they cannot find any comparable industry and country effects within the firm groups?

Shouldn't I care more about the significance of my regressors that the r-squared? And since I clustered the standard errors for firms, isn't it more interesting to add industry and country fixed effects?

[Carlo Lazzaro]

Jan:
it's difficult, for those interested in your thread, to reply without seeing not only what you typed, but what Stata gave you back, too (the FAQ shows you how to share it. Thanks).

[Jan Geurst]

Thank you!

In the attachment you can find my stata output for the different fixed effects.

[Carlo Lazzaro]

Jan:
the FAQ highlight that the best way to share what you typed and what Stata gave you back is via CODE delimiters.
At the top of that, nobody I know on this list (and me with them) would open a spreadshert coming from a potentially unsafe source due to the risk of downloading malicious active contents.

[Jan Geurst]

	Independent variables	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)
	Bond Debt / Assets	Bond Debt / Assets	Bond Debt / Assets	Bond Debt / Assets	Bond Debt / Assets	Debt / Assets	Debt / Assets	Bond Debt / Debt	Bond Debt / Debt
	Panel C
	Eligible x Announcement	0.0068	0.0124	0.0124	0.0124	0.0975***	-0.0124	-0.0114	0.0066	0.1143***
		(0.41)	(0.84)	(0.83)	(0.84)	(3.16)	(-1.12)	(-0.55)	(0.52)	(4.27)
	Announcement	-0.0081	-0.0081	-0.0081	-0.0081	-0.025	0.0158	0.0186	-0.0049	-0.0291***
		(-0.53)	(-0.53)	(-0.53)	(-0.53)	(-1.50)	(1.21)	(1.38)	(-0.50)	(-2.80)
	Ln Assets		-0.0123	-0.0123	-0.0123	-0.0356**	0.0127	-0.0195	-0.0034	-0.0281*
			(-0.43)	(-0.43)	(-0.43)	(-2.21)	(0.38)	(-3.71)	(-0.19)	(-1.77)
	Roa		-0.8770*	-0.8770	-0.8770*	-3.4406***	-0.8759*	-2.6698	-0.1044	-1.0329
			(-1.66)	(-1.64)	(-1.66)	(-2.34)	(-1.79)	(-1.98)	(-0.96)	(-1.55)
	Firm Fixed Effects	NO	NO	YES	YES	NO	YES	NO	YES	NO
	Country Fixed Effects	NO	NO	YES	NO	YES	YES	YES	YES	YES
	Industry Fixed Effects	NO	NO	YES	NO	YES	YES	YES	YES	YES
	Observations	3944	3944	3944	3944	3944	3944	3944	3944	3944
	Number of firms	493	493	493	493	493	493	493	493	493
	R-squared	0.0079	0.2680	0.8969	0.8969	0.4429	0.9029	0.5250	0.8820	0.2222

[Jan Geurst]

Thanks for your advice.

Would this be a clear view of my results? Or do you want to see the stata output in stata format?

the commands I used with the columns are:

1. xtreg dvar var1 var2, fe vce(cluster c)
2. xtreg dvar var1 var2 var3 var4,fe vce(clusterc)
3. reghdfe dvar var1 var2 var3 var4, absorb(firm country industry) vce(cluster c)
4. reghdfe dvar var1 var2 var3 var4, absorb(firm) vce(cluster c)
5. reghdfe dvar var1 var2 var3 var4, absorb(country industry) vce(cluster c)

[Amin Sofla]

could you please share the correlation table as well?

[Carlo Lazzaro]

Jan:
thanks for posting your results (but please note that CODE delimiters implies something else. Please see the FAQ. Thanks).
That said, at their face-value R-squared under (3), (4), (6) & (8) give an impression of overfitting.

[Jan Geurst]

corr indus c incorp, means
(obs=3,944)

Variable | Mean Std. Dev. Min Max
-------------+----------------------------------------------------
indus | 30.76268 14.4386 1 56
c | 247 142.3346 1 493
incorp | 6.874239 3.690321 1 16


| indus c incorp
-------------+---------------------------
indus | 1.0000
c | 0.3173 1.0000
incorp | 0.0762 0.3113 1.0000

Where:
- indus = Industry
- c = Firm
- incorp = country

[Jan Geurst]

Thank you for your responses.

Might the overfitting be caused because the fixed effects for firms created too many 'groups'?

[Carlo Lazzaro]

Jan:
I think that you should report -estat vce, corr- outcome to help us with replying to your query (please use CODE delimiters next time).

CODE delimiters is the #-shaped button in the Advanced editor toolbar. Thanks

[Jan Geurst]

estat vce, corr

Correlation matrix of coefficients of reghdfe model

        e(V) | announ~e  announ~t  l_lnas~s     l_roa
-------------+----------------------------------------
announceme~e |   1.0000                              
announcement |  -0.7465    1.0000                    
  l_lnassets |  -0.6078    0.3981    1.0000          
       l_roa |  -0.0015   -0.3686    0.4791    1.0000

Would this help you?

[Carlo Lazzaro]

Jan:
at a first glance, I would consider omitting -announcement- or -announcement_eligible- from the regression model, as they're highly correlated.