Fixed effect vs random effect vs OLS

[priya sawaliya]

Dear Sir,

I have yearly panel data from 2008-2018 and 63 companies. My variables are:
Dependent- excess return of securities
Independent- KZ i.e index for financial constraints, RD is R&D expenditure to total assets, interaction term i.e. KZ*R&D, log of market capitalization and book to market ratio. All series are stationary.

I first run random effect model and test for xttest0 which suggest OLS. But hausman test suggest fixed effect model.
After this I found heteroscedasticity and cross sectional dependency in my data with the help of xtcsd, pesaran abs and xttest3 command. So I run fixed effect model with vce (robust) this model give four out of five variables significant. My question is:

• Should I go with fixed effect with vce (robust) model? I have attached stata output file below. Please help me .

Thank you
Priya

[Carlo Lazzaro]

Priya:
the accepted way on this forum to share what you typed and what Stata gave you back is via CODE delimiters.
You can also share an example/excerpt of your data via -dataex-.
That said, if you imposed clustered robust standard errors, you should test if the -re- specification fits your data via the user-written command -xtoverid-.

[priya sawaliya]

Dear Carlo,
Apologize for sharing file in this way but i didn't know how to share via Code delimiters and dataex. Thanks for your reply. when i run xtoverid command I get the following:

xtoverid
Error - saved RE estimates are degenerate (sigma_u=0) and equivalent to pooled OLS
r(198);

Is it mean that I should go with pooled OLS and also xttest0 command recommend for OLS as I mentioned in #1 post. But if I'll do so my results are not significant. What should I do sir..
Thank You
Priya

[Carlo Lazzaro]

Priya:
you should go pooled OLS.
The absence of evidence of statistical significance in your sample is not (necessarily) evidence of absence of effect in the population from which your sample was drawn (https://www.ncbi.nlm.nih.gov/pubmed/7647644).

[priya sawaliya]

Dear ,
I read from the below link that xttest0 is useless as it only accounts positive serial correlation of the composite error term.Prof. Wooldridge gave various examples as to how it can give wrong results.He also pointed out that if you have differenced your dependent variable,there is negative serial correlation which xttest0 is unable to capture.He also mentioned having fixed effect is not a problem unless it is correlated with Xit but my doubt is how it can not be correlated with xit as in my example, management of one firm is different from the other which definitely influence my independent variables say for example research and development.
https://www.statalist.org/forums/for...chibar2-1-0000

[Carlo Lazzaro]

Priya:
but you also have heteroskedasticity; hence, robust-clustered statndard errors make sense.

[priya sawaliya]

Carlo sir, you suggest that I should run pooled OLS with robust-clustered statndard errors or Fixed effect model with robust-clustered statndard errors.
Thank you
Priya

[Carlo Lazzaro]

Priya:
you should run pooled OLS with robust-clustered standard errors.
As an aside, please note that Carlo is far enough.

[priya sawaliya]

Thank you Carlo