In article <[EMAIL PROTECTED]>,
Herman Rubin <[EMAIL PROTECTED]> wrote:

Major error, with correction!

>In article <[EMAIL PROTECTED]>,
>nicolexz <[EMAIL PROTECTED]> wrote:
>>I need to sample a random variable from truncated distribution
>>everytime in MCMC.  Suppose, the upper (a) and lower (b) bounds are
>>far from the location parameter (c) and the scale parameter is
>>relatively small, i.e,. b<=a<=c, and c is far greater than a and b. 
>>The chance to sample using slice sampling is trivival.  It's almost
>>impossible to sample it from such a truncated distribution.  However,
>>the problem is quite often in some cases.  What should I do in dealing
>>with this scenario?

<Consider sampling from the truncated normal (0,1) distribution,
<truncated at u and v.  A non-optimal method is to use the same
<truncation points on a distribution whose density is f(x)/3,
<where 

<       f(x) = 1,               -1<=x<=1
*       f(x) = exp(2(x+1)),         x<=-1
*       f(x) = exp(-2(x-1)),             1<=x

<and accept if a test exponential variable T satisfies
<T - ln(f(x)) - (x^2)/2 > 0, saving T if it does, and
<rejecting otherwise.  Truncated procedures can be based
<on this.  However, if the range is outside the interval
<(-2,2), this is not too good.  If the range is u to v,
<where 2 < u < v, a crude method is to let x - u have
<a truncated exponential distribution with density 
<u*exp(-uz), and accept if T - (z^2)/2 > 0.

-- 
This address is for information only.  I do not claim that these views
are those of the Statistics Department or of Purdue University.
Herman Rubin, Department of Statistics, Purdue University
[EMAIL PROTECTED]         Phone: (765)494-6054   FAX: (765)494-0558
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to