[Limdep Nlogit List] negative IIA test statistic

[A N]

Do you mean that I should take away two rows from the permutation matrix J2?
If I got it well, I should alter the procedure as described in Applied
Choice Analysis - why would it be inappropriate?
I can not take away rows unless I do the same with the first matrix, because
to substract the resulting B and V matrices they need to be of the same
dimensionality. After estimating the unrestricted model with six parameters
not associated with constants I need a permutation matrix of six rows, and a
similar one after estimating the restricted model. As the restricted model
is estimated without one of the alternatives, there need to be rows of zeros
in the second permutation matrix (two of the parameters are not estimated).
To strip them off would require also stripping off the rows corresponding to
those parameters from the first permutation matrix, which extracts relevant
elements from B and VARB for the unrestricted model.
In fact all works well when doing it this way, but in doing so we are not
considering parameter estimates of one of the alternatives from the
unrestricted model either, which is not in line with the idea behind the
Hausman test.

In addition, the procedure as described in Applied Choice Analysis (with the
rows of zeros in J2) works well when I exclude the first of the alternatives
that I specified. Even though the resulting matrix vn does have negative
diagonal elements, and so does vninv. But with any other alternative
excluded  I get a negative test statistic...
I've tried all but the one I can't remove because there would be variables
that become constant for the remaining alternatives.

That's why I suppose the problem lies elsewhere, probably in the way I
exclude alternatives - but this relates to the way I organised the data and
I'll need to work it out myself.
I'm only courious on what grounds should the procedure described in Applied
Choice Analysis be altered as you suggested?
Thanks for help anyway!

Aleks