[R] ordered probit or logit / recursive regression

[Ajay Shah]

On Sat, Jun 12, 2004 at 08:11:20PM +0100, Prof Brian Ripley wrote:
> On Sat, 12 Jun 2004, Thomas Lumley wrote:
> 
> > On Sat, 12 Jun 2004, Ajay Shah wrote:
> > 
> > > R is very fancy. You won't get mundane things like ordered probit off
> > > the shelf. 
> 
> But you do (and if you have tried to write them you will discover they are
> not `mundane' to get right).  Two people have already answered for the
> ordered logit, and if you look in the list archives you will find (well,
> you should find) a modification of polr to popr (which will appear in the
> next release of the VR bundle).

I was kidding. I love R. I have been waiting for R for one-third of my
life. I admire R and the R community. I am steadily switching all my
work to R, and my goal is to go to 100% R.

I know the ordered probit model. The ordered probit model is a dear
friend of mine. I have written the ordered probit model in fortran and
in C. I'm quite aware of the difficulties in involved in writing
it:
   We need to estimate tau, the vector of cutoffs, but we can't search
     in R^n since the cutoffs might fall out of sorted order, so we
     use terms like log(tau[2]-tau[1]) as the free parameters in doing
     an unrestricted search.
  Then comes either the delta method or simulations to recover the
     distribution of the untransformed parameters from the MVN
     distribution of the transformed parameters used in estimation.
  And, there are very interesting logic puzzles in identification.

Some years ago, I used to think that clarity on identification and
estimation of the ordered probit model was one of the coolest things
in my brain.

I was trying to be funny and say that "R is so cool; you don't have a
vanilla ordered probit, but you have markov chain monte carlo
inference for an ordered probit". Now _that_ is very fancy (atleast,
in my book). I've lusted after markov chain monte carlo many times,
but never quite done it. Is there a child's guide to MCMC on the net
that I can consume?

> > (I will be very happy if someone will show how to use glm()
> > > to do a vanilla probit!)
> > 
> Quoth Thomas Lumley:
> > glm(y~x+z, family=binomial(probit))
> > Be happy,
> Be even happier that ordered probit also is available.

:-) Thanks! (I haven't seen popr yet).

-- 
Ajay Shah                                                   Consultant
ajayshah at mayin.org                      Department of Economic Affairs
http://www.mayin.org/ajayshah           Ministry of Finance, New Delhi