Multicollinearity

[bibimiwi]

Hello,

I want to estimate a Probit-Modell but I found out, that 2 of my explanatory variables are correlated by 85%.
Those two explanatory variables are:

Dummy1
X1*Dummy1

Does this mean, that I must leave one of them out, or is there any other way?


Thank you very much!

[bibimiwi]

I do have an additional question.

I wanted to test several dummy variables and interaction dummies apon their correlation. In the correlation matrix I found out, that there exist high correlations in the following cases:


Dummy1*X1 with Dummy1*X2
Dummy1*X1 with Dummy1*X3
Dummy1*X2 with Dummy1*X3

Dummy2*X1 with Dummy2*X2
Dummy2*X1 with Dummy2*X3
Dummy2*X2 with Dummy2*X3

Is it possible, that a correlation matrix does not work when using dummy variables???

[startz]

Just run the probit. Multicollinearity correction is built in.

[bibimiwi]

Thank you very much!

Do you know if it is usual to have the dummy variable and the interaction dummy in the same regression model ?
For exampel if I have the following explanatory variables:

X1
X2
X3
D1
D2
X1*D1
X2*D1

Or is it usual to leavy D1 and D2 as singel explanatory variables out?

[startz]

Thank you very much!

Do you know if it is usual to have the dummy variable and the interaction dummy in the same regression model ?
...
Or is it usual to leavy D1 and D2 as singel explanatory variables out?

If you have the interaction, you almost always include the dummy as well.

[bibimiwi]

Ok that helps!

But now I do have the case, where I left all of my original explanatory variables in and only added my interactions. Eviews can't run the regression and gives me the Error Message: "Triangular matrix too small or source Matrix asymmetric". If I leave the dummy coefficients out and only include the intercations, eviews is able to run the regression.
I don't understand the meaning of this error message. Why does the message disappera, when I exclude my dummies?
This message did not appear as I performed THE SAME Regression just on a different dependent variable.

[startz]

You should check whether the different dependent variables have different sets of valid observations. Past that, you might want to post your EViews workfile and the exact commands you enter.

[bibimiwi]

You should check whether the different dependent variables have different sets of valid observations.

What do you mean by different sets of valid obervations?

[startz]

EViews drops NAs. One dependent variable might have different non-missing observations.

[bibimiwi]

I am sorry that I can't upload my workfiles, but they are simply to huge!

I made sure, that no obeservation is missing! But I figured out that my one dummy variable is only represented twice among over 400 other observations. Can this be the problem? But if it is, I wondern why eviews is able to run the regression if I merely include the dummy OR die interaction, but can't run the regression if I include them both,

[bibimiwi]

I do have a compeltely other question..
Is a coefficient of -3.18E-10 with a highly significant p-value worth mentioning? I am not sure beacuse the coefficient is so small, but on the other hand this coefficient does not represent a marginal effect...

[startz]

This depends entirely on the units of measurement of the dependent and independent variables.

[bibimiwi]

the dependet variable is a dummy and the explanatory variable is measured in Million Euro

[startz]

Then the coefficient is probably too small to matter. Having said that, it's odd to get a significant coefficient that's so small. It's possible you have something else wrong with your model.

[bibimiwi]

Can this be the case if a sample selecion bias occured? Because I do estimate an inverse Mill-Ratio of -0.09 at a 10% significance level