I have encountered that whenever deciding upon the model for panel data, it is suggested to perform the Hausman test first and the Breusch-Pagan Lagrange multiplier (LM) test should be performed only if the Random Effect model is suggested by Hausman Test. Could anyone elaborate on why performing BP test is only relevant if the Hausman test suggests RE model and is useless otherwise?
1
-
$\begingroup$ @DimitriyV.Masterov $\endgroup$– AnnaCommented May 16, 2021 at 18:30
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
4
Briefly, the Hausmann Test checks the specification of the model. In particular, you can use it to test if you should be using the RE or FE specification. The BP Test is a check of heteroskedasticity (does the variance depend on the independent variables). You can certainly use the BP Test independent of the Hausman test for any linear model. However, if the Hausman test tells you that you should be using RE then the BP test can tell you wether or not there is heteroskedasticity at the individual level. If none is found then you can just use pooled regression. Basically, the BP test is testing if the individual error term has variance zero, if it does then pooled regression will capture it.
Here is a reference for more details: https://www.princeton.edu/~otorres/Panel101.pdf
-
$\begingroup$ thanks for your reply. When I run the Breusch-Pagan Lagrange multiplier (LM), it says pooled OLS is preferred. Yet, according to Hausman Test, the Fixed Effect model is preferred. Which model I then should use and why? $\endgroup$– AnnaCommented May 16, 2021 at 18:42
-
$\begingroup$ Hey Anna, I think in that case you should go with the Fixed-Effects model since the BP test is really just testing whether or not there is heteroskedasticity. That said, you may need to use this test again in the FE model to ensure you are using the correct standard error specification. This answer might be relevant: stats.stackexchange.com/questions/97023/… $\endgroup$– ArielCommented May 16, 2021 at 18:47
-
1$\begingroup$ thanks! I found a similar question to mine, I have exactly this situation as explained here (link below). From the answer in this post, it is clear that whenever the null of the BP test is not rejected, that means that constant terms are not individual specific and consequently no heteroskedasticity. So, given this, why would one want to use the Fixed Effects model which states that intercepts are individual-specific? Link to a similar question: stats.stackexchange.com/questions/213540/… $\endgroup$– AnnaCommented May 16, 2021 at 20:14
-
$\begingroup$ Okay, so I think if you are saying the BP does not reject the null so that the variance of individual effects is 0, you may be ok using pooled. I think that there is no issue using BP first going back to your original question. The Hausman test which specification between RE and FE to use. $\endgroup$– ArielCommented May 16, 2021 at 21:48