FMOLS: troubles with dummy & negative R-squared

[esterodr]

Hi,

I'm having some troubles with an FMOLS regression, so I will be grateful if someone can help me.

I have an unbalanced panel with 29 countries from 1970-2011 (yearly series).

I want to interact one variable (top1) with the state of development of each country. I've tried two differents ways.


1) The way that I would like to do it.

1.1) I created a dummy (advanced) which equals 1 for advanced countries and 0 for emerging countries.
1.2) I created the variable top1_ad = top1*advanced
1.3) I run this regression, which is saved under the name "regression1":

ca saldo_fiscal dep_joven dep_mayor pib_relativo2 pnae cred_ban cred_oif top1 top1_ad

Method: FMOLS
Panel-Method: Grouped
Trend-specification: Constant
Additional-trends: None

1.4) I received the following error message: "Positive or non-negative argument to function expected".


2) The second-best way to do it.

2.1) Instead of creating a dummy, I used the gdp per capita variable (gdp).
2.2) I created the variable top1_gdp = top1*gdp
2.3) I run the same regression, which is saved with the name "regression2"
2.4) The regression seems to work, but the R-squared is negative: -484.6356


My two questions are:

1) Why I couldn't use the dummy variable?

2) What does the negative R-squared mean? Is the second regression ok or should I drop it?


Thanks a lot for your help.

[EViews Glenn]

Quick answers.

1. TOP1_AD has, by construction, variables for which there is no time-variation within a cross-section. The coefficients on this variable are not identified in the grouped model.

2. The negative R2 is a function of the fact that you are averaging a bunch of estimates across cross-sections. This estimate turns out to have a poorer fit than just using the mean value. There's no reason to believe that the average of the cross-section estimators will have a particularly good fit, not to mention the fact that the FMOLS estimator is not really concerned with sum-of-squares fit as it uses adjustments for the long-run variances and covariances. In my mind, the R2 is really an apples and oranges comparison, but you'll have to decide whether you care or not.

More to the point however, I'm not certain what the meaning of cointegration is in the interactive variable formulation you are adopting.

[esterodr]

Thanks!!! That quick answer helps me a lot!!