Second referee (received 31 January 2000)

==================================================================

Report of the 2nd referee

The paper presents a Bayesian analysis of experimental data with
particular application to the recent results on
\epsilon\prime/\epsilon.

In cases where several results that appear not to be in good
agreement need to be combined to obtain an estimated mean and PDF, a
number of issues come up.  These include, are the errors correctly
estimated, are the measurements truly taken from the same parent
distribution, and are there unknown systematic effects that would
lead to a result being properly rejected or revised?

The author's choice is to modify the uncertainties from the Gaussian
assumption given by the experimentalists to an unsymmetrical
distribution with 2 selectable parameters.  The result for the
combined data aligns with his expectations as a sceptical experienced
physicist.

This is a somewhat idiosyncratic procedure.  Whether the
experimentalists would concur is unclear.  The paper is filled with
generalizations that seem hard to document.  For example, the
Bayesian approach is characterized as one that appeals to the
intuition of most physicists, but certainly not to many that this
reviewer knows.  The use of Bayesian methods has come to the fore as
an approach to handling results from experiments in which there are
physically excluded regions for the parameter being measured.   Most
physicists are actually quite uncomfortable with the concept of a
prior other than zero.  Instead, it is easier to address the
occurrence of unlikely-looking distributions in two steps.  First, if
this is indeed a statistical fluctuation, what is the best estimate
for the true value and for the width of the distribution?  Second,
what is the probability of obtaining the measured values given the
estimated widths provided by the experimenters?  When the latter
probability is low, sceptical experienced physicists usually note
that fact in the form of a reservation about the consensus result.
The addition of extra arbitrary parameters is not very helpful or
useful.

These separate things (and the author's derivation too) do not
actually require appeal to Bayesian methods, although the mathematics
are simplified thereby at the cost of much generality.

The response of the author to the matters raised by the previous
referee is disappointing.  In no case does he offer a constructive or
quantitative answer, but instead simply charges the referee with
"bias" and "prejudice".  Particularly disappointing is the response
to the question, what is the motivation for the choice of
distribution (the gamma function).  The author simply says it is
clearly described in the paper.  This referee too searched in vain
for this description, so it is not clearly described.

The paper contains nothing that will illuminate the long-standing
problem of handling discrepant results, and contains furthermore a
number of unsupported and misleading generalizations that have
negative pedagogical value.  Publication is not warranted.
================================================================