why do I get negative Rsquared in GMM?

[ceycey]

I need some help on the following estimation. I am using an unbalanced panel of european firms where I give cross section fixed effects to countries. I am using Gmm and using 1 period lagged independent variables as instruments.
My problem is I am getting negative R-squares! why does this happen? how do I interpret them? as for certain other samples where I am using the same estimation, I am not having any problems.

Please help!

best regards,
ceycey

Here is the output:

Dependent Variable: TLTA
Method: Panel Generalized Method of Moments
Date: 07/27/10 Time: 10:30
Sample: 1 33623 IF EMU=1
Periods included: 9335
Cross-sections included: 3
Total panel (unbalanced) observations: 11373
2SLS instrument weighting matrix
White cross-section standard errors & covariance (d.f. corrected)
Instrument list: C INT(-1) GRRGDP(-1) DMGDP(-1) SMTO(-1) FFGDP(-1) NFATA NEWSIZE PROF

Variable Coefficient Std. Error t-Statistic Prob.

C 3.061544 1.615473 1.895137 0.0581
INT -29.27661 17.95112 -1.630907 0.1029
GRRGDP 0.048518 0.161835 0.299796 0.7643
DMGDP -0.439105 0.345106 -1.272376 0.2033
SMTO -0.696454 0.376036 -1.852094 0.0640
FFGDP 1.431276 1.600090 0.894497 0.3711
NFATA -0.030261 0.012723 -2.378416 0.0174
NEWSIZE 0.023504 0.001675 14.02940 0.0000
PROF -0.027423 0.008144 -3.367231 0.0008

Effects Specification

Cross-section fixed (dummy variables)

R-squared -1.360221 Mean dependent var 0.530096
Adjusted R-squared -1.362299 S.D. dependent var 0.199136
S.E. of regression 0.306067 Sum squared resid 1064.359
Durbin-Watson stat 1.313768 J-statistic 2.90E-16
Instrument rank 11.000000

[samijo]

In using IV estimation, the R-squared is no more bounded between 0 and 1. In general there is no measure of Goodness of Fit in IV except checking for the validity of the instruments.

[ceycey]

Hi there,

Many thanks. The validity of instruments I believe are ok, so How do I do the goodness of fit check in IV?

[samijo]

Formally, there isn't! But I would take the square of the correlation between the original and fitted values as a pseudo R-Squared!

[ceycey]

ok, thanks.