Attribute Gage RR

Six Sigma – iSixSigma Forums Old Forums General Attribute Gage RR

This topic contains 12 replies, has 7 voices, and was last updated by  Manuel Rosado 14 years, 3 months ago.

Viewing 13 posts - 1 through 13 (of 13 total)
• Author
Posts
• #39949

Allison
Participant

What are the consequences of not having an equal number of  “good” and “bad” parts when doing a Gage R&R?

0
#122776

AT
Participant

The purpose of conducting Attribute gage R & R is to ensure that good parts are always accepted and bad parts are always rejected. You establish a K=1.00 for an acceptable value.. That is the reason why you  have 2-3 parts which are known bad parts in the study. I do not know any reason/logic why one should have equal number of good and bad parts in the study. Perhaps you can explain where have you heard this?
Thanks,
Regards,
Ramesh

0
#122788

Allison
Participant

When I got my BB training by our master black belts, I was told the atttribute Gage R&R would work if   we have minimum 30 parts, minimum 3 operators, minimum 2 repetitions, half of the parts good, half of the parts bad, but never given an explanation on why the gage would not work with unequal parts. This is the way they learned when the company I work for first implemented 6S back in 2001 and hired a company to train the first wave of blackbelts. I  am really curious.

0
#122791

Restagno
Member

The answer is easy Allison, the more known bad parts you have in your study, the probability of getting a positive GR&R result rises. That’s from a probability perspective but I think that the requirement of having a minimum of 30 parts, half good half bad has already some bias in it.
Are you a certified black belt?

0
#122795

Allison
Participant

Thanks you for your reply Sergio. Yes, I am a certified black belt and conducted  gage R&R s based on what I learned in class and understand the relation between number of parts and confidence level etc but can not figure out why the number of bad parts has to exactly match the number of good parts. I am told even a combination of 12+18 would not work. I have not found any explanation in the refernce books I have and would like to fully understand the reasoning behind it.

0
#122808

Cindy
Participant

I always encourage people to have good parts, bad parts and marginal parts.  Most attribute measures are pretty good at telling extremely good from  the really bad — but what you really want to know is how good a job does it do on the marginal.  I pulled out the AIAG MSA manual and it does not specify 50/50…….but it does state that it is desirable that some of the parts are slightly below and above the specification limit,,,,,,,,,,

0
#122819

Dr. Scott
Participant

Allison,
In an Attribute GR&R you want the sample under study to represent what you “suspect” the actual distribution to be. So if it is 50 good and 50 bad, then a 50/50 will tell you something. If it is 98 good and 2 bad (suspected) then you should sample as such.
I also agree with the other post that you should sample across the range (high, medium, low). HOWEVER, if you can do this, then you should use a continuous GR&R type study. The usual.
Hope this helps,
Dr. Scott

0
#122831

Allison
Participant

Thanks for your response and Cindy’s. Since nobody so far has heard about a firm 50/50 rule, I will take this to my certified MBB. I always try to go with marginal samples and stay away from gross defects. Changing it to variable Gage R&R is not easy, since in most cases I am looking at paint defects on car and truck bumpers.
Allison

0
#122848

Darth
Participant

Sorry Scooter, I can’t quite agree on the surface of what has been presented.  We need to establish what the characteristic is that is being tested before declaring that they use a continuous GR&R.  No problem doing an attribute study for other than good/bad characteristics.  We can also do them for nominal and ordinal data as well.  Kappa works fine.  I also agree with the poster that said it might be a good idea to select good, bad and marginal to determine the sensitivity of the measurement system in identifying borderline items.

0
#122851

Dr. Scott
Participant

Sorry Darty. But what exactly do you disagree with. The proper sampling plan for a GR&R, or the fact that if you can identify more than two levels then you can probably identify even more levels (e.g. making it more continous-like) with a bit more effort.
Regards,
Dr. Scott

0
#122853

Darth
Participant

Dr. Scooter, sorry for the absence of the proper salutation.  I guess in a perfect world you can assume that but the poster followed up with an explanation that she was classifying types of defects.  You can probably stretch and encourage her to convert to some sort of continuous measure but I get a sense that she was describing a nominal application of a discrete GR&R.  She might even get to the point of assessing some sort of ordinal scale by which she can classify the defect from 1-5.  That still leaves it in the attribute world and not the continuous.  While the continuous GR&R is sexier, I get concerned about the misapplication.  I had that recently with some BBs who had been coached by someone prior to me.  The characteristic was obviously discrete in nature but she was encouraged to use the continuous version of GR&R.  The final results didn’t make much sense.  Good luck next week and be sure to try the Outback although it will likely eat up your per diem….no pun intended.

0
#122854

Dr. Scott
Participant

Darth,
No worries here. Many call me that around here.
I did not see the email that explained that the poster was talking about types of defects. Only that she was trying to judge good and bad.
I would not try to “stretch” nominal measures into continuous. Perhaps I would ordinal, in the right case. Stretching ordinal to continuous is a very viable and proper thing, if done right. Treating ALL ordinal as attribute can be worse than treating SOME ordinal as continous.
Thanks for you good wishes.
Dr. Scott

0
#123528