EViews.com

Dear all

I am estimating a panel data (country: 56, Year :19) for which there are some variables are non- time varying such as distance and language.

I understand from econometrics textbooks and earlier posting in Eviews Forum that Fixed Effect model cannot estimate non –time varying variables.

However under the panel option, when I selected cross section-none, period- fixed, I was able to estimate the non-time varying variables. But if I selected cross section-fixed, period-none, I will get a message, near singular matrix.


I was puzzling what Eviews did so that we can estimate the non time varying variables under period –fixed but choose none for cross section? I was reading the Panel Estimation method in Chapter 37 of Eviews User Guide. But it was not explain there. Could anyone helps?

I thank you in advance

soo

For period effects models you can't have non cross-section varying variables. For cross-section effects models you can't have non time-varying variables.

Dear Glenn

Thanks for your reply!

I forgot to mention in my earlier post that I have unbalance data (some data are missing!). My understaning is that once the data unbalance, we can't choose both effects at the same time. In other words, I have to choose either / or when the data is unbalance. Am I right?

soo

I know we do the unbalanced two-ways fixed effects. There are a few models that we don't do because the shortcut algorithms that we employ don't work in the unbalanced case, but I know we spent some time getting the two-way fixed effects model to work in the unbalanced case. In the case of the models that don't work you can always add dummy variables yourself for one of the effects (taking care to avoid the dummy variables trap).

Dear Glenn

noted! thanks for your prompt reply..


regards

soo

Hello,

Sorry to get into your question but I am having a similar problem.

I am doing a panel data again. And I am having the same problem as before.

I m using Eview 7.

I am picking up from Excel the panel data

The columns are: Year/Quarter (eg 2001q1, 2001q2), Firm, EBITDA... etc... i have dummies for the years and the quarters

Im using the "basic structure" Panel Data with "cross section ID: Firm" and data series: "Year/Quarter"

I want to know that when I am estimating using LS and I go to Panel Options and put Cross-Section "Fixed" and Period "Fixed" on effects specifications i do not need to include the dummies for year and quarters.

Best regards and thanks in advance

Gaston

When you click on the Proc/ Structure/Resize Curent Page in the current workfile, what does it say?

When you click on the Proc/ Structure/Resize Curent Page in the current workfile, what does it say?

Hello Glenn

I have attached a frame of what appears when I do that..

. What do you think? There is no need for the dummies, right?

Thanks for your help

Does your YEAR variable indicate Year and Quarter? The built-in period effects will be for whatever is defined by YEAR.

Does your YEAR variable indicate Year and Quarter? The built-in period effects will be for whatever is defined by YEAR.

Glenn

Hello

Thanks for your help.

Yes. Year is both.

2009Q1
2009Q2
2009Q3
(...)
2011Q4

so I do not need the dummies... Am I right?

I put fixed effects on period and cross section and the regression gives me the coeficients that I am looking for (the other combinations with random or none, do not). But the Durbin Watson is at 0,6. An the R2 is too high 0,97.

I am replicating an OLS regression from a paper that i read with different data.

The regression is: log(dependant)iyq = constant + b1 (Market Share)iyq + b2(cancellation rate)iyq + ay + aq + random disturbances.
i = firm
y = year
q = quarter

and bc and by are vectors containing year(y) and quarter effects(q)

I am more a time series person. I have not done Panel Data for too many years.

Any help?

Thanks in advance

You're fine on the period effects. They'll be at the quarter level. If you go to the View menu you'll see an entry Fixed/Random Effects, where you can look at the estimated values.

I wouldn't really look at the DW stat. You're better off doing an LM style test (we don't do it automatically, but it's not that hard to do yourself).

As to the R2, it's calculated relative to the constant only model (not the effects only model) so you'll get a high one if your dummies are significant.

You're fine on the period effects. They'll be at the quarter level. If you go to the View menu you'll see an entry Fixed/Random Effects, where you can look at the estimated values.

I wouldn't really look at the DW stat. You're better off doing an LM style test (we don't do it automatically, but it's not that hard to do yourself).

As to the R2, it's calculated relative to the constant only model (not the effects only model) so you'll get a high one if your dummies are significant.

Glenn,

Thank you for your guideline.

What do you mean by "calculated relative to the constant only model"?

For the LM test should i used this http://forums.eviews.com/viewtopic.php?t=334&f=4 ?

or to generate a variable from the data from RESID and regress it against the independent variables and Resid (-1)?

And use it R^2 (n-1) against the critical value. If it is more, then reject and say that there is serial correlation.

Right?

Best regards and thanks!

The R2 is calculated using the difference in the SSR between the full model you estimated and a model with a common intercept only.

I'm not certain I understand your question about the LM test, but Gareth's program should compute the right thing.

The R2 is calculated using the difference in the SSR between the full model you estimated and a model with a common intercept only.

I'm not certain I understand your question about the LM test, but Gareth's program should compute the right thing.

Glenn thank for your answer. I thought that you meant other thing when you talked about the R2.

With respect with the LM test i thought if a grab the data that appears in the RESID variable.

I generate a new variable (because I cannot use the RESID one) and regress that new variable against the independent variables and itself lagged one time.
The i do the test R^2 (n-1) against the critical value.

At the moment sounded okey to me... but i could be talking nonsense...

Thanks again

Best regards

Gaston

A further question on the LM test? If I am correct, the program suggested by Gareth is a Brausch-Pagan-LM-test to test for heteroskedasticity and not for serial correlation?
Is it possible to rewrite the program to get a Breusch-Godfrey-LM serial correlation test?