the Importance of R square in Panel Data Analysis

[bok1234]

Dear Mr.Doan,
I (fixed effect analysis with RATS) and my colleague (system GMM with STATA) tested several panel data analyses.
Coefficients of those tests were good but R squares were terrible - 1 to 10 percents.
In case of panel data analysis R squares are also important like usual regression tests?

[TomDoan]

There is almost no situation where the R^2 provides any particularly useful information. In a time series regression, R^2=.99 might be really small. In a cross section regression, .10 could be really "good". If you're comparing two or more models done with the same dependent variable and same data set, the R^2 provides no information that isn't contained in more useful summary statistics like the log likelihood or sum of squared residuals. The only possible value that it really has is if you compare one set of work with one set of data with a similar set of work with a similar set of data where the fact that R^2 is unit-free is helpful. But you have to have something sufficiently similar (same basic model for a different country for instance) for any comparison to be useful. Absent that, it's pretty much a space filler in a paper.

[bok1234]

Thank you, Mr.Doan. I got to learn the new - at least to me - meaning of R^2. I have known the meaning of R^2 as the explanation power of the model (or independent varibles). However, your teaching seems to focus on the 'relative' explanation powers of alternative models as the meaning of R^2.
Would you introduce me your theses or another scholars' books which explain more the meaning of R^2 just like your perspective or explanation?
I am sorry for asking many questions simultaneously at several sub-issues in the RATS forum. This is because of my recent multitasks. Could you answer me one of my recent questions about system GMM?

[TomDoan]

I'm not sure I could point to any text that would be useful for that---quite a few simply barely mention it. (Hamilton, for instance, mentions it in passing on one page in a 700+ page book.) Again, it's one thing if you have something that offers a direct comparison---for instance, Stock and Watson's book has several model groups where they do a number of regressions with differing explanatory variables and provide R^2's showing that the R^2 may go up from .05 to .30 if you add x1, x2 and x3 to show how the first regression is completely inadequate. Now that same information could be gleaned from the sum of squared residuals, but it's probably easier to understand in R^2 form. But that's where the context in which it is being judged is clear.

[bok1234]

Dear Mr.Doan,
Thank you for your teaching.
By the way, I am wondering whether my question about system GMM could be wrong or not because you have never referred to it. Then, I will correct my question sooner or later and then ask it again.