[Limdep Nlogit List] negative IIA test statistic

A N aleks.nnn at gmail.com
Thu Jun 21 07:42:09 EST 2007


Bill,
Vn is not positive definite, indeed.
But no wonder there are negative diagonal elements - it comes from the
substraction Vr-Vu.
the code, as in the book, for my 19 parameters, 6 attributes model is:
MATRIX; J2=[
1,0,0,0,0,0,0,0,0,0, 0,0,0,0,0,0,0,0,0/
0,1,0,0,0,0,0,0,0,0, 0,0,0,0,0,0,0,0,0/
0,0,0,1,0,0,0,0,0,0, 0,0,0,0,0,0,0,0,0/
0,0,0,0,1,0,0,0,0,0, 0,0,0,0,0,0,0,0,0/
0,0,0,0,0,0,0,0,0,0, 0,0,0,0,0,0,0,0,0/
0,0,0,0,0,0,0,0,0,0, 0,0,0,0,0,0,0,0,0] $

MATRIX ; Br=J2*B ; Vr=J2*VARB*J2' $

? so in Vr the last two rows and two columns are all zeros, while in Vu
there are values >0 (there was no elements set to 0 for the unrestricted
model)

MATRIX ; bn= Bu-Br ; vn= Vr-Vu $

? as a consequence, in vn the two last diagonal elements are negative.

MATRIX ; vninv=[vn]$
MATRIX ; list ; q=bn'*vninv*bn$

Sorry but I can't find a mistake there, it seems that in the example in
Applied Choice Analysis (p. 600) there is also a row and a column of zeros
in Vr, so after substracting Vu (which is not shown) one diagonal element of
vn would be negative...
Where am I wrong then?

Aleks



More information about the Limdep mailing list