Skip to content

Intelligent use of mcmcsamp()

1 message · John Maindonald

#
I wonder whether you have considered simulation from the full
variance-covariance matrix.  For some samples, one or more
variance parameter estimates will be allowed to go negative, 
but the full variance-covariance matrix should still be positive 
definite.  This avoids any problem with variance close to zero,
but it does mean (and I think this a useful side-effect) that 
credible intervals will sometimes have a negative lower bound,
or even lie entirely on the negative real line!

Why a useful side effect?  I've encountered cases where scientists, 
in a field experiment, have chosen blocks that are e.g., long strip 
at right angles to a river bank, thus spanning about as wide a range 
of variation as is possible.  (I've been told about the river bank 
example, the examples in my own experience were a bit different.)
A model in which there is a negative component of variance may 
then give a formally correct variance-covariance structure.  It is a bit 
like using a complex number representation to solve some problems 
on the real line.

John Maindonald             email: john.maindonald at anu.edu.au
phone : +61 2 (6125)3473    fax  : +61 2(6125)5549
Centre for Mathematics & Its Applications, Room 1194,
John Dedman Mathematical Sciences Building (Building 27)
Australian National University, Canberra ACT 0200.
http://www.maths.anu.edu.au/~johnm
On 08/12/2009, at 9:30 AM, John Maindonald wrote: