Hashed hash

[Berzerk <berzerk@xmission.xmission.com>]

OK, I have been doing a few numerical experiments on hash functions to 
see if all this stuff I have been saying is true.  I took the folowing 
function, as my n bit to n bit hash function.

first n bits(md5(n bits))

and iterated it to see how many colisions there were.  I found that the 
total entropy in the result typically decresed by 50% for n=8,10,12,14 
and droped like a rock when you itterated these.

I have a couple of questions,
1) is this a good hash function, or am I missing something here.
2) the expected collision rate for rand functions is much lower.  I am at 
a loss to explain md5.

I will be trying smaller versions of all of the suggestions here to see 
if they help or hurt, and will set them up to run on the spare cycles on 
a machene or two around here.

Any comments on my stratigy are appreciated in advance of me running the 
calculatios.

Roger.