RE: den.list-d Digest V2003 #38

["John McConnell" <wysowl@msn.com.au>]

John Dowd wrote,
With regard to independence, I'm going to be lazy and not dig up a bunch of
references, but hasn't Wheeler sown that a fair amount of auto-correlation
(clear lack of independence) can be encountered without materially affecting
the usefulness of the control chart?  Perhaps some additional comments along
this line would be helpful to those who are confused.

There is plenty of evidence to suggest John is correct.  Wheeler's work
provides some great insights.

Right now I am working with metallurgical and cement plants.  Nearly all the
process data lack independence; they auto-correlate to some degree.
Nevertheless, control charts are being routinely and successfully used as a
monitoring tool in reporting systems and as an analysis/monitoring tool in
production processes.  True, autocorrelation of data brings some
limitations, but even supervisors and managers with a two-day introductory
course under their belt and a couple of months experience are managing to
extract much useful information.

The technical types know when to react to a signal in (say) the chemistry,
particle size, density or metallurgy, and when to ignore what is most likely
to be random variation.  The managers know when a dip in the production
figures is random and when it signals an actual fall in production.

At the cement plant they have an instrument that uses a radioactive isotope
to read the chemistry of the raw meal on a conveyor.  This data is highly
auto-correlated and there is a tremendous amount of noise in the raw data.
They had tried moving averages and several other approaches to put this data
to good use.  All of them failed.  However, controlled experiments and some
trial and error revealed a sub-group size that if plotted on an average and
range control chart provided greatly improved chemistry control which led to
better fuel efficiency, higher tonnage output and improved uniformity in the
product.  So, even noisy, auto-correlated data can provide some useful
information and lead to significant improvement.

The test of any theory is that it matches observation.  The observations at
these plants show that control charts work, even with auto-correlated data.
They are not perfect, but neither is any other method.

Cheerio!

John McConnell