(Panel analysis) Breusch-Pagan LM test interpretation

[Changhyeon Song]

Hi, I'm Song


In my model, there is time=invariant variable. So I can't use fixed-effect model.

then I consider which one is more appropriate between pooled OLS versus Random-effect model.

So I tested var(u)=0 (Breusch-Pagan LM test) and I got a result Prob>chibar2 = 1.000

this means that I cannot reject the null, and we cannot say "there exist a random effect"

Q) I suspicious about the result of probability=1.0 (1.0 is so extreme I think)

when I try the same procedure after I remove time-invariant variable, the result are not that extreme. (probability exist between 0 and 1)

Is it wrong with time-invariant variable?

Then can I use Breusch-Pagan LM test with time-invariant variable?


Sincerely

I always get help from this wonderful website. Thank you all the time


PS) I know that there are some discussions about this topic in STATA LIST

https://www.statalist.org/forums/for...chibar2-1-0000

https://www.stata.com/statalist/arch.../msg00011.html

https://www.statalist.org/forums/for...=1516605239077

[Carlo Lazzaro]

Changhyeon:
presence of time invariant predictors does not imply, per se, that you cannot use -fe-; your choice should (also) depend on the -hausman- test outcome (that you do not report).
As per your post, -xttest0- does not support the evidence of a random effect in your panel dataset: however, as stated above, this is not enough to rule out -fe- specification from your methodological toolkit.

[Changhyeon Song]

Thank you, Carlo

But I have interest in coefficient of time-invariant variable (also with time-variant variables)

So if I choose -fe-, I cannot figure out the coefficient of that variable.

[Carlo Lazzaro]

Song:
you might be interested in taking a look at: https://blog.stata.com/2015/10/29/fi...dlak-approach/