Professor David Finney has asked me to submit the following query to
allstat. Please send replies to me, not to the list.
Chris Theobald
School of Mathematics, University of Edinburgh and Biomathematics &
Statistics Scotland
[log in to unmask]
__________________________________________________________
In the official instructions for an important medically related
calibration procedure, I read: "Any samples with a perpendicular
distance greater than 3 residual standard deviations from the regression
line should be excluded. After removal of such samples, the final
orthogonal regression line should be calculated".
This rule is stated without any reference to a textbook or other source,
but possibly with an implication that it is analogous to a standard
practice in estimating regression equations.
I am involved only peripherally and I am arguing against adoption of
such a simplistic rule. My urgent need is to locate any textbook or
manual of laboratory practice, in which something like this rule is
stated - either dogmatically or with explanation. I would like an
authoritative source, but even a bad though popular book would be better
than nothing!
I'll be interested in any replies. In this context, if there is to be
any rule for removing outliers (possibly none is really needed), it must
be simple and capable of incorporation into a simple software program.
It will be used in many places as a routine operation, by trained
technicians but with no involvement of, or supervision by, professional
statisticians.
|