Residual tests in panel estimations

[jma]

Dear all

can anybody tell me why I only have the possibility to make a "Histogram Normality Test" when I structure my workfile in panel format and run a random or fixed effects regression?

Somehow eviews does not offer me in this case access to the full range of tests provided in a pooled format (Correlogram Q-statistics, Heteroskedasticity tests, Serial Correlation LM test etc.)...

Thank you!

Jörn

[EViews Gareth]

I don't think that is quite true - Pooled equations don't offer those tests either (as far as I know).

The reason that panel type tests are not offered is simply that they are a (sometimes a lot) more complicated than the non-panel equivalents, and we have not yet had time to implement them.

Some of the tests are probably (at least in a simple balanced panel case) along the lines of auxiliary regression tests, which can be done manually.

[jma]

Gareth,

thanks for your reply. Maybe my wording was a bit misleading: with pooled format I meant the case of an unstructured workfile, i.e. when you throw all observations in one "pool".

It would be great if you could clarify two other things for me:

- As I have written, after estimating an equation based on panel data there do not exist any residual tests in the equation toolbar (View/Residual tests...). However, the estimation output window does report a Durbin-Watson statistic. Additionally, when I save the residuals as a series, I am able to obtain a correlogram & Q-stat
- Now the question: Do these tests account for the panel type data, i.e. is the serial correlation tested within the cross-section units across the periods? Because otherwise, these statistics would not help me, right?!

Thanks!

Jörn

[EViews Gareth]

Depends on what you mean by "account for panel type data". The tests will not let lags cross seams, i.e. a lag of the first observation in cross-section 2 is "NA", not the last observation of cross-section 1. In that sense they are accounting for the panel type data.

[jma]

The question for me is whether this is a groupwise test.

E.g., if the Durban-Watson test is 2,0 and the correlogram/q-stat is ok, can I infer that the errors for a particular unit at one time are unrelated to errors for that unit at all other times (or the period before in case of Durbin/Watson...)?

Thanks
Jörn

[EViews Gareth]

They treat the data as one long time series, with the exception that seams are not crossed (as outlined above). Thus I guess you could say it is not a group-wise test.

[jma]

Ok. Does "one long time series" mean that the test averages observations per year, i.e. it estimates the value of a common AR(1) parameter using observations per period of the data set?

[EViews Gareth]

one long time series means that if you have a 100x10 panel, then the time series has 1,000 observations (not accounting for missing values due to lags). No averaging occurs.

[Thea]

As an input calculating test statistics manually, does anyone know how to generate fitted values from running period fixed effects using a pool object?

I'm sure there must be an obvious solution here, but I've been stuck for a week!

My dataset consists of 14 firms, over 3 years with two variables. (They entered the pool from being stacked in Excel according to cross-sections).

I've tried putting in 'pool.fit' as a command (but does not work) and I'm stuck as to how I get Eviews to only apply the coefficients to particular years of data (vectors or matrixes of the coefficients generated cannot then be combined with the original pool variables).

Many thanks for any assistance.

Thea.

[EViews Glenn]

I'm not certain that I understand entirely the part about applying coefficients to particular years of data and the like, but you should be able to get fitted values from a pool in two steps:

Code: Select all

pool01.makeresid r? pool01.genr fit? = y? - r?

[Thea]

Glenn,

Thank you - that does it! :D A bit of innovative thinking and remembering that r = y-yhat were what was required. I appreciate your help.

Thea.

[durbin-watson]

Hi,

I would like to ask three simple questions:

1) I know that the Durbin-Watson stat is not valid as an indicator of autocorrelation when there is a dependent lagged variable on the right side of the equation. But is it true that the DW stat is not valid for panel data IN ANY CASE (ie. before adding lagged variables?)

2) I am told that including the lagged dependent variable is a way to "get rid of" autocorrelation. Is it possible/normal that after adding it there is still autocorrelation in the sample? What do I do then, should I run the Cochrane-Orcutt procedure?

3) Is it possible that after running the Cochrane-Orcutt procedure for order 1 autocorrelation (typing ar[1] at the end of the regression command) there is still a degree of autocorrelation in the estimations? If yes, does it mean that the autocorrelation is of an order other than 1? How do I check this?

Would be very grateful if anyone could answer these questions, project is due in a week! Cheers

[nobodyelse79]

Hi,

I have ran a panel regression and I'm working with different pools. I want to test for heteroskedasticity. From the forum here I seem to understand that no tests exist in E-views to do this in a panel data framework. So how do we do it in practice? I have tried to save the residuals and then run a regression with the squared residuals as dep variable and the regressors on the right hand side.
But when I generate residuals it gives me the option to form a group, but not to save residuals as a series in the pool. I need to save the residuals of three different estimations (with three different dependent variables, but same regressors). How can I do that? If I freeze the residuals as a group, the residuals are not recognized as a series and therefore I can't run any regression. I think I am making some mistakes but I hope you can help me in spotting it.

Thank you for your help.Michele

[ranaamar]

please tell me that DW is valid or not in panel data

[mariowijaya]

i have a problem with my view/residual diagnostic. my eviews7 cant show me white heteroscedasticity, serial correlation LM test, and etc. is the problem on my regression test? or bcause the program is just a demo? cause my view/residual diagnostic only show me histogram - normality test. thank you