Hi,
I'm trying to use GLLAMM to estimate an ordered probit model on a
database with
two levels (i.e. survey of 16 activities in 196 firms). The linear
component with (3)
random effects is as follow:
y_ij=X_ijB + W_ij \theta_i + \epsilon_ij,
where \epsilon is a standard normal variable, and \theta_i is a 3X1
vector distributed
normally accross firms (indexed by i), with mean 0 and covariance
matrix \Sigma (I
want to estimate the correlation betwen the \theta's).
I tried to estimate this model using the following command:
eq: var1=w1;
eq: var2=w2;
eq: var3=w3;
gllamm y X, link(oprobit) fam(gaussian) eqs(var1 var2 var3) nrf(3)
adapt;
However the program failled to complete the first iteration and
reported the following error message:
could not calculate numerical derivatives
flat or discontinuous region encountered
(error occurred in ML computation)
(use trace option and check correctness of initial model)
I tried to run the same model using the binomial probit link (i.e.
collapsing the ordered choice in two choices), and it worked perfectly.
I have also noticed in the GLLAMM manual that the ordered response
examples are always estimated
with the binomial family (i.e. distribution assumption for the random
effects). I tried to use the binomial family instead, but it
wouldn't let me estimate three correlated random effects. Also, I
tried the example from the GLLAMM manual with the gaussian family
instead of the binomial family, and stata reported the same error
message (which indicates that my data-set is not the source of the
problem).
Does anybody have any idea how to estimate this model in GLLAMM ?
Thank you very much in advance,
J-F
