fixed effects and clustering standard errors - dated panel

[zeca]

Hi,
I'm working with a sample of the 100 largest publicly traded banks in Europe from 1996 to 2011 (a total of 1133 observations). Banks are from 25 diferent countries.
I'm using eviews to estimate a regression in which Leverage is the dependent variable and then I have 5 explanatory variables (one of them is a dummy - dividend payer). Please see attached the equation.
I'm using a dated panel data (annual freq, banks are the cross-section) and LS Method. I think is what best fits...

Q1:
Following the available literature in corporate finance I have to use time and country fixed effects (Ct and Cc). Regardind time fixed effects I found it easy (panel options / Effects specification / Period:Fixed).
Could you please give me some guidelines to deal with country fixed effects? (One of my available variables is country)

Q2:
what should I do to adjust the standard errors for clustering at the bank level (cross-section)? Should I use the white period coef covariance method in the panel options? Is it enough?

Q3:
variable Div is a dummy (1 if bank pays dividends in a given period, 0 otherwise). Once is a dummy it has the same processing and interpretation of remaining variables?

Many thanks in advance!
zeca

[EViews Glenn]

Q1. @expand(country, @dropfirst)
Q2. That does robust clustered standard errors. Whether it is enough of not, only you can say.
Q3. For almost all purposes, are the same as any other variable.

[zeca]

Q1. @expand(country, @dropfirst)
Q2. That does robust clustered standard errors. Whether it is enough of not, only you can say.
Q3. For almost all purposes, are the same as any other variable.

Thanks for quickly reply and help!

Q1: just to confirm, is it?
book_lev c mtb(-1) profit(-1) size_ln(-1) coll(-1) dividend @expand(country, @dropfirst)

Q2: I wanted to ask you if it would be necessary to choose a GLS Weights or if the standard errors adjustment for clustering at the cross-section level is done simply by using the white period coef covariance method?

Many thanks!!

[EViews Glenn]

Q1. If I'm understanding what you are doing correctly, yes.
Q2. Whether to do GLS or have standard errors that are robust to departures is really a personal choice. The White period computes the standard errors clustered on cross-section ID.

[zeca]

thanks :D

[zeca]

Q1. @expand(country, @dropfirst)
Q2. That does robust clustered standard errors. Whether it is enough of not, only you can say.
Q3. For almost all purposes, are the same as any other variable.

I'm back because against all my expectations, when I run this equation eviews give me the error "near singular matrix".
book_lev c profit(-1) size_ln(-1) dividend @expand(country, @dropfirst)

There are any other way to take into account the country fixed effects in this case?
I'm sending attached the workfile. Could you please help me with this?

[EViews Glenn]

You're doing fixed effects on the crossid (banks?). Once you remove bank variation, there's no variation in the country. Another way of saying this is that your bankid and countryid variables are perfectly correlated.

[zeca]

You're doing fixed effects on the crossid (banks?). Once you remove bank variation, there's no variation in the country. Another way of saying this is that your bankid and countryid variables are perfectly correlated.

In fact i´m not doing fixed effects on the crossid but on period. But even without consider any other fixed effect, the error msg appears.
you can test by yourself writing: book_lev c profit(-1) size_ln(-1) dividend @expand(country, @dropfirst) in the workfile I posted.
And i do not find evidence that bankid and countryid variables are perfectly correlated...
could you test please to confirm if it works with you?
thanks a lot

[EViews Glenn]

As to the latter..

show @varsby(country,crossid)

shows that there is no variation in the country indicator once I control for the cross-section ID. Another way of saying that is that banks don't change countries in your sample.

I assumed that you were doing cross-section effects. Let me look at the period effects issue.

[edit]

Looks as though there is simply multicollinearity between regressors. I'll leave it to you to find out why. But if you take the spec and estimate with the dummies and remove DIVIDEND, you get estimates...

[zeca]

Hi,
returning to this case after a break of some days, I have an additional question.
In my equation I have to introduce a dummy (lets say "dividend") that is based on the "cash dividend paid". Dividend = 0 if the bank does not pay dividends and 1 if "cash dividend paid" >0 (the bank pays a certain amount of dividends).
Which is the correct way to deal with this?
Introduce into the equation a variable previously built that takes only the values ​​0 and 1, or use the variable "cash dividend paid" and an eviews function to create a dummy?
Many thanks!

[EViews Glenn]

I'm not sure I understand. Is this previously defined variable that you are talking about the same as creating the dummy from the variable "cash dividend paid"?

[zeca]

I'm not sure I understand. Is this previously defined variable that you are talking about the same as creating the dummy from the variable "cash dividend paid"?

To remember this case:
I'm using eviews to estimate a regression in which Leverage is the dependent variable and then I have 5 explanatory variables (one of them is a dummy - dividend payer).I'm using a dated panel data (annual freq, banks are the cross-section) and LS Method.

In the sample I have the field called "cash_dividend_paid" (million eur). Before I import the data to the eviews, I transformed in excel this field in a dummy variable called "dividend" (0 if cahs dividend = 0 and 1 otherwise.
Then I wrote the following equation: book_lev c mtb(-1) profit(-1) size_ln(-1) coll(-1) dividend

My question is:
Instead of import and use directly the binary variable "dividend" in this equation, may I import the variable "cash_dividend_paid" and transform it in a dummy just in the equation with the @expand or whatever...?
The result is the same?

many thanks

[EViews Glenn]

If the transformation you are doing in EViews is the same as the one in Excel, of course. Arguably, most of the transformations you'll want to do are more easily done in EViews.

You should be able to code the dummy as "cash_dividend>0".

[yzerman]

I'm back because against all my expectations, when I run this equation eviews give me the error "near singular matrix".
book_lev c profit(-1) size_ln(-1) dividend @expand(country, @dropfirst)

There are any other way to take into account the country fixed effects in this case?
I'm sending attached the workfile. Could you please help me with this?

zeca,
I get the same error as you in a very similar situation in my own panel data. I discovered that ­@expand(country) includes a dummy for every country in your database, even if some countries are excluded from the estimation sample. For example, if you set your sample to exclude banks from Germany (e.g. if country<>"Germany"), the dummy for Germany will still appear in your equation for the non-German observations in the form of a vector of zeros.

Alternatively, if you have missing values for several variables in your equation that result in the German observations being dropped from the estimation, you will also get a vector of zeros appearing in the place of the German dummy. Either way, a vector of zeros causes an error.

If this is your problem, you would need to create another version of the country variable that includes only the countries actually represented in the estimation sample. This new variable could then be used with @expand.

[shajjudu]

Hi,
If data set comes from one country and one industry, For example, data from Australia and banking industry (say, 20 banks) for 6 years (2007-2012) Do I still need clustering data? can I do that through Eviews package? If so please give me the guideline. Many thanks