***  For details on how to be removed from this list visit the  ***
***          CCP4 home page http://www.ccp4.ac.uk         ***


On 8/31/05 4:45 AM, "[EMAIL PROTECTED]" <[EMAIL PROTECTED]> a écrit :
 
>> You are also using a non-traditional definition of 'statistical bias'. The
>> Gauss-Markov theorem guarantees that if condition (1) above holds [even if
>> (2) and (3) are violated], then the least-squares solution is unbiased. Here
>> I am using the standard def of 'bias' in statistics:
> 
> I'm using the same definition - more or less ;o) Only from a Bayesian
> viewpoint, what corresponds to a 'point estimate' would be the expectation
> value of the posterior probability for which you, of course, need the
> distribution.

Well, I can see how there is a Bayesian analog of 'bias' (though I have
never seen a precise def), but it is still a very different concept from the
frequentist usage of 'bias', which is meaningless for a Bayesian.

> If you integrate over a different distribution, you'll often get a
> different estimate. If you only _use_ the mean and the variance, you are
> in effect committing yourself to a normal distribution.
> 
> Ok. We have to separate two issues. One is the set of assumptions laid out
> by a traditional LS derivation and the other is the 'biased view' that
> Bayesian Theory is the only reasonable scientific inference machine. You
> are right that LS does not explicitly make any assumptions about the error
> distribution, however if you analyse standard ad hoc traditional
> techniques from the Information Theory and Bayesian point of view (which
> can be done with some care) you often reveal the real hidden underlying
> assumptions on which you base your inference.

OK -- I was not intending to defend the frequentist position (I actually
care little about bias), but rather simply explain the traditional
justification for LS in terms of the G-M Theorem, which does not depend on a
Gaussian error distribution.

However, and this is really a subtle point, I still disagree that a Bayesian
analysis reveals a hidden normality assumption for least squares. From a
Bayesian and/or Info Theoretic POV, if you have only the first two moments,
you are best off acting as if the distribution is normal. That is different
from *assuming* that it is normal. If your analysis is based on a normality
assumption, it is valid only to the extent that the normality assumption is
valid. In contrast, a Bayesian is saying that it doesn't matter what the
real distribution is, normal or not, you are best off using a normal
distribution in your calculations, because a normal distribution assumes the
least structure (or information) in your data given only the first two
moments. Use of any other distribution would involve extra unwarranted
assumptions. 



Reply via email to