Dummy variable in a Fixed effects (FE) model

#1
Hello :)

I am doing a paper where I examine the Scandinavian welfare model effect on inequality in the Scandinavian countries. I examine it using a balanced panel data, where I use a dummy variable for countries that use the Scandinavian welfare model.

I have used the Hausman test to find out that I need to use a fixed effects model (see the attached image). But my problem is, that I can not use a dummy variable in a fixed effects model. So I can not examine the Scandinavian welfare model effects by using my dummy variabel.

How do I solve this problem?

Looking forward to hearing from you.
 

kiton

New Member
#2
Hello!
What prevents you from including the dummies in a FE model? You can use any variables in the FE model specification, as long as these variables are time-variant.
 
#3
Hello!
What prevents you from including the dummies in a FE model? You can use any variables in the FE model specification, as long as these variables are time-variant.
FE longitudinal models assume that there are no time invariant predictors (i.e. things like age and ethnicity).

Now if you are including time invariant predictors, you need to have a random effects model, Hausman test be ****ed :). It does not matter if your errors terms are not correlated in a model without the dummy variables. You want to use dummy variables which most certainly will make your error terms be correlated. In which case, you need to use the random effects model.

Remember, let your research questions guide your methods, not the other way around.
 

kiton

New Member
#4
To add to the replies above:
It's not actually about the Hausman test alone to decide on the model's effects. Rather, its about the assumptions about the unobserved confounders.

In case of the author's needs, a time-invariant dummy estimated via RE specification will till be biased anyway.
 
#5
To add to the replies above:
It's not actually about the Hausman test alone to decide on the model's effects. Rather, its about the assumptions about the unobserved confounders.

In case of the author's needs, a time-invariant dummy estimated via RE specification will be biased anyway.
Also, to note, the Hausman test assumes you do not know the structure of your error correlation matrix. If you know that you will have a time invariant predictor within your model, then there is no reason to run it.

But to be honest, I cannot think of an example of a time invariant dummy variable, because by its very nature, its shared across subjects (i.e. 1 and 0). That is unless you have only two subjects :).