Re: Use of Control Charts

[DANSWART@aol.com]

Dear Mr. Bevan,

I am not a control chart expert and cannot speak to the subtle mathematical 
issues you speak of.  However, I believe you have a real opportunity here to 
guide the team focus away from specific chart methodology to process 
methodology.  (How many can say that their company statistical expert has 
read Shewhart, Wheeler, et al. is interested in hearing competing views on 
control charting?)  The key IMO is not the chart methodology you choose, it 
will be a human judgment in any case.  The key is getting the team to 
understand the inductive inference inherent in the chart, and accepting that 
the ONLY test you can perform to validate it is EMPIRICAL.  In other words, 
the choice you make now must be examined in light of future experience and 
should not be the last word on the subject.

My recommendation is:

1.  Strive to get the team to commit to a scientific process of testing the 
operation of statistical control no matter the methodology selected for 
selecting the chart and setting the chart limits now.  Get them into the PDSA 
cycle with respect to the operation of statistical control, without spending 
all your time, energy, and credibility defending a specific chart 
methodology.  Shewhart says on p. 30 of SMFTVOQC (see below) the following 
with respect to your situation:

"Obviously, there is no a priori, formal, and mathematical method of setting 
up a criterion that will indicate an assignable cause in any given case.  
Instead, THE ONLY WAY ONE CAN JUSTIFY THE USE OF ANY CRITERION IS THROUGH 
EXTENSIVE EXPERIENCE (emphasis added)."

I believe here Shewhart means your experience as well as others.  In short, 
pick a method, try it and see if it works.  What does it mean to "see if it 
works?"  It means ONLY that "...if assignable causes are looked for when an 
observed statistic goes outside its control limits such causes are ALMOST 
ALWAYS FOUND (emphasis added)."  SMFTVOQC p. 33

If the chart constructed under the chosen method reliably detects VERIFIED 
assignable causes, IT WORKS.  This is what I understand to be the difference 
between analytic and enumerative studies: no amount of math theory will 
validate your methodology in an analytic study, only future data.

You may even be able to keep parallel charts [say XmR charts] to compare 
results.

[Caveat: Shewhart tells us that the chart batting average will not be 1.000 
under any conditions.  It will likely detect many verifiable signals at 
first, then diminish in accuracy.  "My own experience has been that in the 
early stages of any attempt at control of a quality characteristic, 
assignable causes are always present even though the production operation has 
been repeated under presumably the same essential conditions.  As these 
assignable causes are found and eliminated, the variation in quality 
gradually approaches a state of statistical control as indicated by the 
statistics of successive samples falling within their control limits, except 
in rare instances, and by the fact that when assignable causes are looked for 
in these rare instances they are seldom discovered."  SMFTVOQC p. 37]

I recommend that the team commit to reviewing their SPC work using the 5 
steps of SPC found in SMFTVOQC p. 25, and the 4 practical requirements 
Shewhart felt were necessary for effective control charting found in SMFTVOQC 
p. 30.

If the team accepts these as worthy of consideration a periodic review is 
possible as part of a PDSA cycle.

Thanks for your question, it got me back into reading Shewhart again.

Note:  SMFTVOQC = Statistical Method from the Viewpoint of Quality Control by 
Walter A. Shewhart (ISBN 0-486-65232-7)


Dan
Dan Swart, CPA, APC
2720-21st Street
Sacramento, CA  95818-3129

danswart@aol.com
(916) 451-1040     local voice
(916) 451-1047     fax
(888) 706-1040     toll-free
========================================================================