multicollinearity concerns panel data across section or with

[Wouter van der Stee]

Dear,

By generating a correlation matrix between the independent variables of my panel of data I checked for possible multicollinearity problems. I have 4 different sections. My question now is whether I should look at the correlations between the variables only within an section or also across sections to check for possible multicollinearity problems.

Thanks in advance!

Wouter van der Stee

[startz]

What multicollinearity "problems" are you checking for?

[Wouter van der Stee]

If the explanatory variables are strongly correlated this can lead to inflated standard errors leading to unstable results. So I guess just for the regular multicollinearity problems.

[startz]

If the explanatory variables are strongly correlated this can lead to inflated standard errors leading to unstable results. So I guess just for the regular multicollinearity problems.

What you say is true, but often misunderstood. The standard errors may be larger than one wishes, but they are still correct. Multicollinearity simply means that there is not much information in the data about the separate effects of the variables.

[Wouter van der Stee]

Oke thank you! I recently read another post of yours about serial correlation and GARCH(1,1) models.
If my residuals contain serial correlation and an ARCH test shows I should use a GARCH(1,1) model, how can i correct for serial correlation in the GARCH model? I do not find such an option in the option menu...

[startz]

I think you just include ar(1) in the mean equation, but I haven't tried it.

[Wouter van der Stee]

ok thanks!

[Wouter van der Stee]

startz,

What is the difference between including an ar(1) or a lagged dependent variable in the equation?

sincerely

Wouter

[startz]

The equation

Code: Select all

ls y x ar(1)

is equivalent to

Code: Select all

ls y = c(1)*(x-c(2)*x(-1)) + c(2)*y(-1)

AR(1) and a lagged dependent variable are only equivalent if there are no right hand side variables.

[Wouter van der Stee]

So which one is better to include in the mean equation in order to control for autocorrelation?

[startz]

ar(1)

That's exactly what the ar(1) command is for.

[Wouter van der Stee]

ok thank you very much!

[Wouter van der Stee]

Would it be wrong to use a ldv instead of ar(1)? The results (coefficients etc) vary dramatically, but both seem to remove autocorrelation

[startz]

If by "ldv" you mean including dummy variables, then yes. If that gets rid of the serial correlation then it's probably a better solution than using ar(1)

[Wouter van der Stee]

I mean lagged dependent variable so y(-1). So what youre saying with your last sentence is both are valid to do but lagged dependet variable is better? Can it be determined which is best to ( lagged depenedent or ar(1)) do since the results vary quite substantially?