On Wed, 13 Oct 2004, Ignacio Díaz-Emparanza wrote:
I have been building some examples about heteroskedasticity data
with gretl. I had two variables whose linear relationship had
hetero. and so I regressed y on x using 1/sqrt(x) as weigths.
Then, I suspected that the residuals were heteroskedastic yet, so
I tried to do some testing on them.
Firstly, I cannot do a LM test on the WLS regression (¿Why not?
umm I have to think a while about it ... I know that the
regressor now is xt*wt, and so the test may have a different
interpretation, but ... may have any sense in this case, ¿or
I'll have to think about that too.
Secondly, as I did not have the test, I saved the WLS residuals an
regressed them over a constant (by OLS), so I had a window model,
and the heteroskedasticity test was active in the corresponding
menu. I selected this and ... gretl exited!!
I found the bug responsible for the crash. Fixed in CVS.
In fact I have seen that regressing any variable over a constant,
the same occurs. I understand that the Hetero. test runs a
regression of the residuals over the regressors and their squares,
That's right, and so naturally it fails in this case.
and so in this case there were no regressors, but I think Gretl
should act in another form. For example in this case gretl may ask
for a variable, or use the index ...
I would day that in this case (regression on only a constant), gretl
should make the heteroskedasticity test menu item unavailable. I
tend to think it is the researcher's responsibility to save the
residuals (which are just de-meaned y) and then regress them on
whatever variable might be relevant.
Department of Economics
Wake Forest University, NC