Glen wrote:
> [EMAIL PROTECTED] (David C. Howell) wrote in message
> news:<[EMAIL PROTECTED]>...
>> My daughter just asked me a question that I should be able to
>> answer, but
>> can't. Any help would be appreciated.
>>
>> It is well know that the sum of normally distributed random
>> variables is
>> itself normally distributed. But I want to know the sampling
>> distribution
>> of the sum of lognormal distributions.
>
> It's not anything that has a name, or even a nice closed form.
>
>> My first thought would be that the central limit theorem would
>> suggest that
>> the sampling distribution would approach normal as N increases.
>
> True - when the CLT applies, it applies. But don't expect it to come
> in very quickly!
>
>> On the
>> other hand, when I generate multiple independent lognormal
>> distributions
>> and then take their sum, a P-P plot tells me that the sum is nicely
>> lognormal, not normal.
>
> You have rediscovered something many, many people have noticed
before.
> I have seen it called the permanence of the lognormal - the tendency
> of sums of lognormal random variables to have an approximately
> lognormal distribution.  (I've seen it mentioned in papers in
> geostatistics and in optometry - or some similar eye-related
research
> area, perhaps opthamology - for example.)
>
> In many other application areas where the use of the lognormal is
> common, it doesn't have a name but people have still noticed it. I
> think it's one of those things you tend to hear about by word of
mouth
> a lot more than by reading it in papers.
>
> I've seen it often (it's relevant to an application I've been
involved
> with) - and it happens quite nicely even when the lognormals aren't
> independent. I've dealt with cases which include several /hundred/
> correlated lognormals (the underlying normals being linearly
> correlated), and in the cases we deal with - perhaps 99% of the
time -
> the lognormal was an extremely good fit to the sum of lognormals
> (substantially better than gamma, and much, much better than
normal),
> even right into the tail (looking at p-p plots)
>
> (NB: This is not inconsistent with CLT - lognormals with small sigma
> can look pretty normalish, so a sum of lognormals can continue to be
> approximately lognormal while it's approaching normality.)
>
> I have played around with some algebra and found out when it's not
> going to work so well, but those circumstances don't tend to occur
in
> the applications we deal with.
>
> Glen

The usual book "Continuous Univariate Distributions" vol 1 (Johnson,
Kotz & Balakrishnan, 1994) has a short discussion of this, with
references to a few publications, but there seems to be nothing much
extra that can be done in this loose context.

David Jones


.
.
=================================================================
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