Hi, Simon.
No bug here. Look at the model for ordered probit in the documentation.
If the cut points are theta and the parameters are tau, then theta2 = theta1
+ exp(tau2). Therefore, the parameters (tau) do not need to be ordered
because they are constrained by the model parameterization to be greater
than the previous cut point (which is achieved by the exp).
Yours,
Olivia
On Wed, Oct 12, 2011 at 3:55 PM, Simon Hug <simon.hug(a)sunrise.ch> wrote:
Dear Zelig
I came across something a bit disturbing in the ordered probit model of
Zelig (version 3.4-5). More specifically the simulations carried out after
a model estimation
generate coefficients (in $par) for the cutoff-points which are not
correctly ordered for all simulation runs. This obviously makes it
impossible to check the predicted probabilities generated and stored in
qi$ev.
If I do the exactly same thing with oprobit.bayes the gamma-coefficients
(as they are called there) are for each simulation run in the correct
order and the predicted probabilities in qi$ev correspond to what one
would generate with the help of the coefficients in $par.
What is wrong here?
best wishes, simon