denton

There are a couple of approaches.
One is to simply dump all the data from all three batches into one pile, and do the familiar calculations on all the data, using subgroup size 1.  Cpk will be calculated using the moving range method, and Ppk will be calculated with the sum of squares method.  
The second works if you can consider each batch as…[Read more]

Just run Capability (Binomial) in Minitab.  It reports a Z score for the process.

Autoliv supplies air bags, seat belts, etc. to the automotive industry, and has a stellar Six Sigma program.

For situations with 3-4 input variables, a full factorial is the economical answer.  When you get to 5 inputs, half fractionals become attractive.  At 6 or more inputs, P-B designs can be the best alternative.
The best approach that I have seen is in Wheeler’s Understanding Industrial Experimentation.  He takes an approach that is different fr…[Read more]

Mann-Whitney requires that both samples have the same distribution.
If your sample size is large, you can usually almost ignore the normality requirement for the T test.
Minitab will compensate quite nicely for unequal variances.  If the variances are very unequal, it will reduce your effective sample size, but give you the correct answer.  J…[Read more]

Control limits are derived only from the data.  Specifications and tolerances do not enter into the calculation.
A control chart detects change in the process.  A process can be in control, and also be reliably unsatisfactory.  If you want to know if the process is satisfactory, use a Capability Study.

It’s an interesting problem. 
To solve it, I think you will have to take pairs of samples, a millisecond or so apart, and assume that nothing inside the pipe changed in that amount of time.  You are then justified in saying that most of the within sample variation is from the measurement system.
Take 25-30 pairs of data, and perform Int…[Read more]

Almost all real-world data will test non-normal if you get enough of it.  As your sample size increases, you can detect ever more minute deviations from normal.
Normality is generally overrated, anyway. :-)
With 522 samples, T tests aren’t going to care if the data are normal, and it won’t matter for control charts.  You might get into some d…[Read more]

There are a couple of different models.  There is the less often used “attribute” and “variable” data model, and there is the “nominal”, “ordinal”, “interval”, and “ratio” model which is much more widely accepted.
“Continuous” and “discrete” are subsets of “ratio” and “interval”.  Data are interval if they can meaningfully be added and sub…[Read more]

If you have population data, then tools like T tests are not needed, and normality is irrelevant.  You can just take the mean of your population, or the difference in means between two populations, and that’s your answer.
Tools like T tests estimate population parameters, based on samples.  They estimate the probability of being wrong in your a…[Read more]

Drop me a note at [email protected], and I’ll send you some material that should help.  I gave a paper at SME on “Staying out of the Hands of Charlatans”, and I’ll give you a copy of the slides.  I also have some other stuff that should help.

Contact me at [email protected].  I can put you in contact with someone who can probably give you a name.

Six Sigma is a customer centered, systematic, data driven method for doing things better.

Wheeler did extensive work on the issue of normality and control charts, and wrote a dandy little book on the subject.  He clearly demonstrates that normality is simply not required for control charts.  Just chart the raw data, whether it is normal or not.  In practice, the difference you get by transformation is small, and the potential for co…[Read more]

The nested GR&R in Minitab will, indeed, do the job.
A somewhat simpler technique is Interclass Correlation Coefficient.   It’s not in Minitab, but it is an excellent technique for these situations.

Best practice is a “can’t tell zone” that is two Effective Resolutions wide, centered on both the USL and LSL.  If something falls in that zone, you acknowledge that you can’t tell whether it is good or bad.  That gives you three output states:  good, bad, and can’t tell.  As you improve your measurement system, the “can’t tell zone” gets sma…[Read more]

One good way to do this is with Interclass Correlation Coefficient.  It is simpler than GR&R, and will probably answer your question.   It is also excellent for gauges where destructive testing is involved.  The basic math is straight out of Xbar and R charts, and works slick as can be.  For some very short help, you can contact me at [email protected].

I really enjoyed the one day I spent in Denton.  Visited the old courthouse.  Nice town.
I suspect that if we could get together over lunch, we would probably really enjoy ourselves, discussing stats.
There is still a problem with your approach.  A-D is usually based on a sample.  For the reasons already outlined, it, too, will fail in the fac…[Read more]

I’m in the wilds of Utah…  though I’ve been to Denton TX.  Thought it was very nice of them to name the town after me with only one visit….  :-)
The problem is establishing normality.  That is only slightly easier than bringing a pint of neutrinos back from the sun.  It takes millions of observations to characterize the tails of the dist…[Read more]

Mean and standard deviation “drag” their uncertainties with them into the Cpk and Ppk equations.  The uncertainty in standard deviation is usually the limiting factor.
Also, all sampling systems fail in the face of good quality.  If your process is running 1,000 defectives per million, a sample of 100 has less than a 10% chance of containing a d…[Read more]