Very low r-squared at xtreg fe

[Tim Meier]

Dear all,

I have a question about my R-squared. The term is very small.

1) Is my model still meaningful or do I need to change the independent variable?
2) I have also already changed all variables as a check and also used alternative dependent variables. R-squared almost always remains at 0. Can the "problem" be due to the data set?

Basically, I proceeded as follows:
After I could not find any significant correlations between my variables, I did the Hausman test with the result that I should calculate fixed effects. Additionally I calculated "xtoverid" with the following result: Error - saved RE estimates are degenerate (sigma_u=0) and equivalent to pooled OLS. Following the statement of this thread (https://www.statalist.org/forums/for...he-same-as-ols), one should then follow the Hausman test.

Afterwards I used the following code for the calculation:

quietly xtreg c.VD NDF c.BB c.EE c.AS c.KR, fe vce(robust)
estimates store reg_0    
quietly xtreg c.VD c.PL NDF c.BB c.EE c.AS c.KR, fe vce(robust)
estimates store reg_1                                
quietly xtreg c.VD c.PL c.GE c.IE NDF c.BB c.EE c.AS c.KR, fe vce(robust)
estimates store reg_2                                
quietly xtreg c.VD c.PL c.GE c.IE  c.PL#c.GENDF c.BB c.EE c.AS c.KR, fe vce(robust)  
estimates store reg_3                                
quietly xtreg c.VD c.PL c.GE c.IE c.PL#c.IENDF c.BB c.EE c.AS c.KR, fe vce(robust)
estimates store reg_4                                
quietly xtreg c.VD c.PL c.GE c.IE c.PL#c.GEc.PL#c.IENDF c.BB c.EE c.AS c.KR, fe vce(robust)
estimates store reg_5    
estout reg_0 reg_1 reg_2 reg_3 reg_4 reg_5 using VD_Reg.csv, nobaselevels cells(b(star fmt(3) label(Coef.)) se(par fmt(2) label(std.errors))) varwidth(23) starl(+ 0.1 * 0.05 ** 0.01 *** 0.001) eform stats(N df_m ll chi2 r2_a, labels ("No. of obs." "Degrees of freedom" "Log likelihood" "R2") fmt(3 0 1)) label legend varlabels(_cons Constant)

The regression table looks as follows:

Thanks for your advice and help!

[Carlo Lazzaro]

Tim:
1) r-sq: if you go -xtreg,fe- is the within R-sq the one you should consider;
2) the outcome of the community-contributed module -xtoverid- (that BTW focuses on -re- specification) is telling you that your dataset does not show evidence of a panel-wise effect.

[Tim Meier]

Hi Carlo,
thank you very much for your reply.
1) I used the within R-sq.
2) Does this mean that I should not calculate a panel but pooled ols?

[Carlo Lazzaro]

Tim:
1) have you already check that your regression is not misspecified?
2) if there's no evidence of a panel-wise effect, you should switch to a pooled OLS (with standard errors clustered on -panelid- as observations within panels are not independent).

[Tim Meier]

1) I checked the model specification using the procedure in this thread: https://www.statalist.org/forums/for...d-effect-model

predict yhat, xb

gen yhat2 = yhat^2

gen yhat3 = yhat^3

gen yhat4 = yhat^4

xtreg c.VD c.PL c.GE c.IE  NDF c.BB c.EE c.AS c.KR yhat2 yhat3 yhat4, fe vce(robust)

test yhat2=yhat3=yhat4=0

( 1) yhat2 - yhat3 = 0
( 2) yhat2 - yhat4 = 0
( 3) yhat2 = 0

F( 3, 13) = 1.32
Prob > F = 0.3112

If I understand the result correctly, there is no misspecification of the model.



2) I'm a little lost. Isn't it generally always more advisable to use the panel instead of the pooled OLS estimator for panel data? At least that's how I understand the discussion starting at #8 in this thread (https://www.statalist.org/forums/for...logit-decision).

[Carlo Lazzaro]

Tim:
is -sigma_u- still zero in your

xtreg c.VD c.PL c.GE c.IE  NDF c.BB c.EE c.AS c.KR, re vce(robust)

?

[Tim Meier]

Carlo:
Yes

[Carlo Lazzaro]

Tim:
if (as it seems from -sigma_u-=0) there's no evidence of a panel-wise effect, you should switch to a pooled OLS (with standard errors clustered on -panelid- as observations within panels are not independent).

[Tim Meier]

If I omit the variables for the test for model misspecification and run the regression with fixed effects, my sigma_u is >0.

 
 xtreg c.VD c.PL c.GE c.IE  NDF c.BB c.EE c.AS c.KR, fe vce(robust)

I thought this sigma_u value shows the existence of a panel wise effect and confirms the use of the fixed effects estimator?

[Carlo Lazzaro]

Tim:
it seems you've switched to -fe-.
The usual recipe is to compare it to -re- option via the community-contributed module -xtoverid-.
As as aside (that comes from FAQ), please use CODE delimiters to report (in full, please) what you typed and what Stata gave you back. Thanks.