#195 Documentation: mention Baltagi/Li (1991) for Breusch-Pagan test

[KTTK]

By playing a bit with gretl's and R's (plm) Breusch-Pagan test, I found out that plmtest (in R) does not treat unbalanced panels while gretl does - at least, I feel like gretl uses the method of Baltagi/Li (1991) for unbalanced panels. I think, the documentation could reflect this. What do you think?

[Also I have prepared a rather rough implementation for R, see http://stackoverflow.com/questions/31988449/implementation-of-breusch-pagan-test-for-random-effects-in-plm-with-unbalanced-p and the GitHub link mentioned there]

[Sven Schreiber]

Sorry for not getting an answer so far. Strictly speaking you're right that this could be considered a documentation bug, so I'm leaving this (just lowering priority a little). However, for discussions like this the gretl mailing list are currently still a better place, see http://lists.wfu.edu/mailman/listinfo/gretl-users.
(I don't know the answer to your question, sorry.)

And BTW, does your R implementation match gretl's results for some test cases?

thanks,
sven

[KTTK]

Thanks for getting at this issue.

My own implementation of the Breusch-Pagan test for unbalanced panels as in Baltagi/Li (1990) [not (1991) as in my original post] matches gretl's test exactly. Thus, I can safely say that gretl uses that. Please update the documentation.

Full reference is:
Baltagi/Li (1990), A lagrange multiplier test for the error components model with incomplete panels, Econometric Reviews, 9, pp. 103-107

[KTTK]

In addition:
There is also a Breusch-Pagan test for heteroskedasticity. Thus, it would be good to distinguish that one from the Breusch-Pagan test for inidvidual effects. The current output in 2015d just calls the test performed "Breusch-Pagan test" (However, advanced users can distinguish based on the null: "Null hypothesis: Variance of the unit-specific error = 0").

[Sven Schreiber]

Yes a remark should be added, in the section "Testing panel models" in chapter 18. Furthermore I've noticed that the "modtest" command has an option "--breusch-pagan" that indeed is about the heteroskedasticity test which I find a little problematic because of the ambiguity you mention.
But at least the documentation fix is easy.
cheers,
sven

[Sven Schreiber]

Here's a link to my post on the mailing list about this:
http://lists.wfu.edu/pipermail/gretl-devel/2016-March/006644.html
cheers,
sven

[Sven Schreiber]

OK the "testing panel models" section is in chapter 19 (as of August 2017), not 18.
Let me pick up this thread by raising a more fundamental question, however: How can one retrieve the test results of the automatically printed tests (Breusch-Pagan idiosyncratic variance etc.) in a script way? The principle in gretl is normally that everything can be done either in the GUI (with some exceptions) or by programming with the Hansl scripting language.
thanks,
sven

[KTTK]

For the record, it is answered here (with the comment that it is yet to be documented how to get an overview of the extra accessors): http://lists.wfu.edu/pipermail/gretl-devel/2017-October/008143.html

Use:
eval $model.bp_test

[Sven Schreiber]

Ok, let me try to summarize before this gets too confusing:
1) Baltagi/Li for the unbalanced case should still be mentioned, only that the chapter of the guide is a moving target, now it seems to be 20 (Nov. 2017). (But the tex source file of course is not affected by the automatic numbering.)
2) The $model accessor inspection could/should be documented (but noting that the built-in reference for $model already says: "additional information may be available ... you can get a copy of the bundle and print its content, as in / ols y 0 x / bundle b = $model / print b").

That's it for this documentation bug I think, the other things have been fixed.

[Sven Schreiber]

I have just edited in git the user guide, mentioning Baltagi-Li and also the relevant $model bundle members. Closing this, thanks for the patience.