GRR – prove significant improvement
Six Sigma – iSixSigma › Forums › Old Forums › General › GRR – prove significant improvement
 This topic has 6 replies, 4 voices, and was last updated 12 years, 10 months ago by Stefan Szemkus.

AuthorPosts

January 16, 2009 at 9:13 am #51691
Stefan SzemkusParticipant@stefan Include @stefan in your post and this person will
be notified via email.Hi everybody,
I read this forum for several moths but this is the first post I do – I hope i do everything correct.
My question:
How can I determine to have achieved a significant improvement of my measurement system when I execute a GRR before and after a change of the measurement system.
E.g.: I did an initial GRR and got 82% (versus Total Variation), then I changed the measurement process and repeated the GRR, using the same parts and appraisers => I achieved 33% GRR.
Although it obviously looks like a significant improvement, I would like to prove that.
Could you help me? How do you prove improvements of GRR results?
Many thanks in advance,
Stefan0January 16, 2009 at 9:48 am #179813Hai Stefan.
Strange question. Generally you want to prove that your measurement system is Ok to ensure that data analysis on the measured data can be trusted (i.e. the data analysis gives outcomes that can be connected to the situation the data describes). So as soon as you are good enough (<10% or 30% whatever your treshold is) then that's it.
But here is what I thought up (if it is wrong another expert will probably correct, but you may have to wait for Monday since some of them are at the i6Ssummit):
You can prove that you changed significantly with the Equal Variances test. But you have to do that not on all the variance but on the measurement variance only. Take the two datasets of the gage r&R investigations. For each dataset seperately do the following:Substract from each datapoint the average measured value of the product.
This way you have neutralised the product variation in your data. All the variation left in the dataset is the variation ’caused’ by repeats and operator change.
Now perform an Equal Variances test on the two datasets. It will hopefully show that there is a significant change.
Good Luck.
Hmm, at rereading i found a possible flaw. With my ‘solution’ you will compare the absolute variances while the gage r&R investigates relative variances. I’ll have to think some more about this.
0January 16, 2009 at 10:02 am #179814Hi StefanIf you remeasure your components and recalculate
your process capability index (Pp or Ppk) you will/
should see and improvement in it – it will increase
– as the contribution to the overall variation
observed made by the measurement system variation
decreases. This will also result in a narrower
tolerance interval. Practically you will see less “good” parts being
rejected in error due to measurement system error
or less escapes. Have you a control chart in place?
You might consider recalculating the control limits
on the charts after a little while. By the sound of things you have made a significant step forward but you still have more to do to
improve the gauge. Best of luck
Eoin0January 16, 2009 at 9:06 pm #179838
Bower ChielParticipant@BowerChiel Include @BowerChiel in your post and this person will
be notified via email.Hi Stefan
William Woodall and Connie Borror published a paper entitled “Some Relationships between Gage R&R Criteria”. It is available at http://www3.interscience.wiley.com/cgibin/fulltext/114280277/PDFSTART. They refer to another paper by Burdick, Borror and Montgomery entitled “Design and Analysis of Gage R&R Studies” which gives formulae for calculating confidence intervals for % Gage R&R. If you calculated 95% confidence intervals before and after and got say 70% to 94% before and 20% to 66% after the fact that these intervals don’t overlap would provide reassurance of a real improvement. Statistical purists would likely object but it might be worth looking at.Best WishesBower Chiel0January 20, 2009 at 11:26 am #179944
Stefan SzemkusParticipant@stefan Include @stefan in your post and this person will
be notified via email.Dear remi,
thank you very much for the quick reply on my question. I was really astonished that you bothered with my problem in that short time.
Would you say it is not necessary to prove the significance on the improvement? I always have been teached to prove those things, but perhaps that is not the way it works in practice.
If I want to execute a 2variancestest, what are the degrees of freedom I have to choose? Could you help me also with this question?
Thank you very much in advance,
Stefan0January 20, 2009 at 11:30 am #179945
Stefan SzemkusParticipant@stefan Include @stefan in your post and this person will
be notified via email.Dear Eoin,
thank you very much for answering my question.
I’m very angry with myself, because the way you proposed to prove significance should have come into my mind too.
Again, thank you very much!
Best regards,
Stefan0January 20, 2009 at 12:13 pm #179947
Stefan SzemkusParticipant@stefan Include @stefan in your post and this person will
be notified via email.Dear Bower Chiel,
thank you very much for the detailled reply. I will get the papers and do the CI calculations.
If I do all three recommended calculations (2variancestest, Cpkcalculation and CIcalculation) that should give evidence enough to prove significance, shouldn’t it.
Best regards,
Stefan0 
AuthorPosts
The forum ‘General’ is closed to new topics and replies.