GLS weights used in forecasts

[rdwest]

Hi,

I am estimating a panel model with cross-section fixed effects using GLS. My expectation is that the residual ought to differ by cross-section (i.e., there is cross-sectional heteroskedasticity). To correct for this, I'm using GLS weights. So my model looks more or less like this:

equation eq_d1.ls(CX=F,WGT=CXDIAG) log(y) log(y(-1)) log(y(-2)) log(y(-12)) LOG(x1) LOG(x2) LOG(x3) @expand(month, @dropfirst)*log(x4) C

Two questions:
1. Is there a formal way to test the hypothesis of cross-sectional heteroskedasticity (against the null of no cross-sectional heteroskedasticity)?
2. My impression is that the GLS weighting procedure just involves calculating weights as the inverse of the variance of each cross-section across the time-span of the data, and applying these in the estimation procedure. (I asked previously whether there is a way to extract the value of the weights to check this, and was told that there is not.) Thus, once the model parameters have been estimated, the weights should not figure into EViews's calculations when I use the command "eq_d1.forecast", right?

Thanks,
R.



UPDATE:

Okay, now I'm truly confused. My data (aviation data) are at a monthly frequency, and my previous hypothesis was that good indicators of this month's value of the dependent variable should be the last several months' values (i.e., last month, and the month before -- so two monthly lags) and the value one year prior (i.e., a twelve-month lag -- since travel has a strong annual cyclical pattern), and a host of exogenous explanatory variables. Thus, I was running the specification shown above. However, when I try to test the hypothesis that the effect of the one-year lag is unimportant (which I can't do, strictly speaking, because removing the lag extends my sample back in time), I get the error message that says "Positive or non-negative argument to function expected in computation of weight group." Testing this, it seems to be the two-month lag that is the problem: If 1) that term is dropped out, and a one- (or other-) month lags are included, or 2) a longer lag (3-month, 12-month, what have you) lag is included in addition to the two-month lag, EViews is apparently able to calculate the weights and run the regression. If not, I get that message. This doesn't make sense to me: If the GLS weights are calculated on the variance of the dependent variable, I fail to see why restricting the sample to begin two months after my initial date (by including a one- and a two-month lag) creates zero-variance problems, but restricting the sample to begin one month after the initial workfile date (by including a one-month lag) does not.

Insight into this, or the above, anyone??