[Limdep Nlogit List] Pages E21-37&38 of Limdep manual 9.0 for Maddala's Model 3

[Richard Lo]

Dear All,


In pages E21-37&38 of Limdep manual 9.0, there is a program code for
Maddala's Model 3 (p244-245).
I have two questions for the code.

1. the calculation of c   "   c = sigmasq1 - 2*gamma1*sigma1*rho12  "
  It seems that in Maddala's book, the estimated variance refers to that of
the first reduced-form equation instead of the structural equation.
Shouldn't 'sigmasq1' and 'sigma1' be replaced by 't1sq' and its counterpart?


2. the definition for G'<v0>G for the calculation of va2
   From the expression "z2xv=z2'xm3*v0i*xm3'z2", it seems to imply that
G=xm3'z2
    However, xm3'z2 does not correspond to the matrix G=[pie1,J2] in
Maddala's book
   (where J2=<xm3'xm3>*xm3'x2 because xm3*J2=X2 ; g=[pie1,J2] , pie1 is the
matrix of
   coefficient estimates for the reduced-form OLS equation). It could be
numerically verified.

In the manual the are two ways to calculate rho12 ( a user written program
in Stata may provide the third way).When these estimates of rho12 differ
widely, do we have any criteria to choose from them?

If we get nonsense estimates of rho12 (e.g. rho12>1 or rho12<-1), does it
still make sense for us to use these nonsense estimates to "correct" the
results?
Thank you in advance for your advice.

Richard Lo