Hi David,
I might be wrong, but I think the fact the theta parameter governs the
equation zeta_i ~ (1/theta) * {something} means theta can't be zero (because
1/0 is undefined) and so has to be reparameterised accordingly for
unconstrained optimisation. (Also, I think by definition, as it's something
one is searching for the ML estimator for, theta isn't an ancillary
parameter.)
http://gking.harvard.edu/zelig/docs/Model20.html
I think it's this reparameterised version (which is valid during the whole
real number line, rather than just the positive part) which is part of the
variance-covariance matrix. One has an approximation of the standard errors
on this from the square root of the last element of the diagonal (from top
left to bottom right) of this matrix. One can de-reparameterise this to get
standard errors on the theta. (slightly confusingly, the var-cov matrix
emerges from differentiating twice a vector comprising the betas then the
reparameterised theta: this 'stack' is conventionally called 'theta',
though
in this case one of the parameters one is trying to find the ML estimator
for is also called 'theta'...)
"how do I get back a correct
variance-cov matrix?"
I think the short answer is that, as long as the reparameterisation of theta
is valid for the whole real number line, then this is the 'correct'
var-covar matrix, as one needs to feed this into the MVnormal later to get
estimates (which take into account fundamental uncertainty) of the
unconstrained versions of the parameters. These estimates can then be
converted into the appropriate mus {e^(X*beta), so constrained across the
positive part of the real number line} and thetas (which, as I said, I think
can't be 0, so have to be constrained accordingly) when simulating for the
quantities of interest. (OK, that's not really a short answer!)
Clear? (probably not... any clearer explanations/corrections welcome)
Best,
Jon
-----Original Message-----
From: gov2001-l-bounces at
lists.fas.harvard.edu
[mailto:gov2001-l-bounces at
lists.fas.harvard.edu] On Behalf Of
landau at
fas.harvard.edu
Sent: 10 April 2007 05:08
To: gov2001-l at
lists.fas.harvard.edu
Subject: [gov2001-l] reparameterization for 2c
Hi all,
It seems like you need to reparameterize the ancillary parameter on part c
in
the Negative Binomial section but not the coefficient parameters (far from
sure
about this though). This works fine for the ancillary parameter itself,
which
you can then easily transform back. But how do I get back a correct
variance-cov matrix? Holger had some code on this for section 4 where he
simulated a parameter, transformed it, and then got CIs. But things seem
more
complex where you have a full variance-cov matrix, and particularly where
you
aren't reparameterizing everything.
Thanks,
David
_______________________________________________
gov2001-l mailing list
gov2001-l at
lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
--
No virus found in this incoming message.
Checked by AVG Free Edition.
Version: 7.5.446 / Virus Database: 269.0.0/754 - Release Date: 09/04/2007
22:59
--
No virus found in this outgoing message.
Checked by AVG Free Edition.
Version: 7.5.446 / Virus Database: 269.0.0/754 - Release Date: 09/04/2007
22:59