On Thu, Jul 31, 2008 at 2:22 PM, Jonathan Hanson
<[email protected]> wrote:
What happens when you estimate 2 and 3 component mixtures without
covariates in the probabilities?
I can get results with no covariates in the probabilities when starting with
a 2 component mixture and then moving to 3 component mixture. Usually, as
soon as I add one covariate in the probabilities, Stata cannot find initial
values, whether there are 2 or three components.
That's odd -- if you specify the value of zero for your covariate,
then this is precisely the model with fixed probabilities. If the
latter converged, then at least for one set of values the likelihood
is computable! And most likely in the vicinity of that model the
likelihood should be well defined. Of course Stata might be making big
jumps with the -ml search- procedure getting out of the domains of
reasonable estimates for which the likelihood is computable. It
depends on the implementation details though, and I am drawing a blank
about how the mixture probabilities were coded -- a multinomial logit,
I would think (Partha?)
A crude way of looking into what -from- is expecting is to supply a
vector of zeroes or ones as the -from- option, and also specify
-iter(0)- option (I hope Partha's code supports transfer of
maximization options; this is good -ml- programming style!). Then the
estimation will conclude without starting, and you can look into the
format of the parameter vector using
mat li e(b)
