Checking for model misspecification with panel data

[Titir Bhattacharya]

Hi Statalist,

For some context: I'm attempting a fixed effects regression on panel data. When I performed the regression with classical standard errors, I got significant coefficients on my variables of interest. However, when I took robust standard errors, significance disappeared.

To check whether I have correct model specification, I followed the method outlined here by Prof. Silva. I used the following code:

areg y x i.year,a(ID) cluster (ID)
predict y_hat  

gen y_h_2=y_hat*y_hat
gen y_h_3=y_h_2*y_hat
gen y_h_4=y_h_3*y_hat

areg y x y_h_2 y_h_3 y_h_4 i.year,a(ID) cluster (ID)
test y_h_2 y_h_3 y_h_4

To which I got the following result:

F(  3,  1924) =    0.97
            Prob > F =    0.4051

Can I take this result to mean the model is correctly specified? Or should I check for higher order powers? Does this mean significance on my coefficients disappeared due to the nature of the data and is unlikely to be corrected by changing model specification?

Any help would be greatly appreciated.

Thanks,

[Carlo Lazzaro]

Titir:
you actually followed the procedure reported in -linktest- entry, that you can safely apply after -areg- (but not after -xtreg-; that's why Joao' advice comes in handy when you are dealing with -xt- commands).
That said, the results of -test- tell you that there's no evidence that your model is misspecified.

[Titir Bhattacharya]

Titir:
you actually followed the procedure reported in -linktest- entry, that you can safely apply after -areg- (but not after -xtreg-; that's why Joao' advice comes in handy when you are dealing with -xt- commands).
That said, the results of -test- tell you that there's no evidence that your model is misspecified.

Thank you Carlo for your response. Given there is no evidence of misspecification, would I be correct in interpreting the disappearing significance to be a result of heteroskedasticity due to nature of data? Can you please suggest what should be the correct way to proceed in this case?

Thanks,

[Jeff Wooldridge]

This result does not imply your model isn’t misspecified. It is a functional form test and it looks like you haven’t missed important nonlinearities. I do have to check whether the fitted values from areg include the fitted dummies. Hopefully not.

RESET is not a test of omitted variables.

[Andrew Musau]

I do have to check whether the fitted values from areg include the fitted dummies. Hopefully not.

Not the default with -predict, xb-. If you want predictions with the fitted dummies, you need to specify the option -xbd-

Options for predict


----+ Main +--------------------------------------------------------------------------------------------------------------------------------------------------------------------
xb, the default, calculates the prediction of xb, the fitted values, by using the average effect of the absorbed variable. Also see xbd below.

xbd calculates xb + d_absorbvar, which are the fitted values including the individual effects of the absorbed variable.