"Maarten Speekenbrink" <[EMAIL PROTECTED]> wrote in message news:<[EMAIL PROTECTED]>...
> Oops, I've made a mistake in my former post. the corrected message is:
> 
> Hi there,
> 
> I would like to know the distribution of a linear combination of (not
> identical) logistically distributed variables, i.e.
> 
> f(x) = exp( (x - alpha)/beta )/(1+exp( (x-alpha)/beta ))^2 .
> 
> Does anybody now what this distribution would look like, or know a good
> reference? I've looked in the 'Handbook of the logistic distribution' by
> Balakrishnan, as well as in Johnson & Kotz, but couldn't find anything.
> Since the distribution is very similar to the normal, I would expect a
> similar result to that for the normal distributed (it is logistically
> distributed itself). Thanks for your help.

As Herman has already said, it won't be logistic.

Let's call the variables X1, X2, ....

You want the distribution of a1 X1 + a2 X2 + ... + an Xn.

Note that if Yi = ai Xi, this is the same as the distribution of 
the sum of the Yi. The Yi are of course logistic as well (with 
parameters that are easily computed).

Under certain conditions (e.g. constraining the relative variances
of the Y's so that a small fraction of the variances don't dominate; 
that there not be too much dependence) the central limit theorem 
should begin to kick in once you have a big number of terms.
(That is, if the conditions hold and you have enough terms in 
the sum, the distribution will start getting close to normal.)

Clearly the distribution will be symmetric. Herman has already
noted that it will be log-concave.

I don't know if bounds will help you, but bounds could be computed.
An easy bound is to ignore the extra information in it being 
log-concave and just work with the fact that it is symmetric and 
unimodal.

This leads to a chebyshev-like inequality that is tighter than
the usual chebyshev (with 1-1/k^2 replaced by 1-4/9k^2, if memory
serves - so 44.4% of the tail area you get in the usual chebyshev). 
This will give an upper limit on proportions in the tail.

To be handwavy for a moment, if the Y's are not too dissimilar 
in variance, the tails should tend to be lighter than the logistic. 
Without doing some analysis that's probably beyond me at the moment, 
this would suggest that under some circumstances the logistic would 
also be an upper bound on tail area. But since I can't give you a 
good idea of how wide a range of cicumstances that would happen 
under**, we can't really call it anything more than a vague 
approximation which may often tend to be larger.

** of course for specific situations you could do some 
computations and compare (or even some simulations if 
that suffices for your purpose).

The normal would seem (again without doing any analysis to back 
it up, so again don't take it seriously unless you have something 
more to go on than my hand-waving) to provide a lower bound on 
tail areas.

There are a few circumstances where even bounds of "symmetric 
unimodal" and "normal" might be helpful.

Precise answers would probably require doing some form of 
numerical convolution for each specific circumstance.

Glen
.
.
=================================================================
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