Although I have seen a similar thread somewhere before, I still have the urge to ask because I still have not found a satisfying answer that can quench my thirst.

Why on earth we one needs 10 samples to do a GRR? Isn't having too many parts will introduce a component variability that is due to the process that was used to make the 10 parts. I heard, somewhere, that there are 2 types of GRR studies. One is Gage only RR and the other is Gage/process RR. if you are interested in the 2nd one then yeah sure go and choose 10 parts but if we are only interested in validating the gage, then shouldn't we minimize the # of parts. I would think that ideally one should use only one part or 10 "identical" parts if they are available.

Lastly, lets say that we have 2 sets of reading (say 1- 4 parts measured twice each and 2- 4 Parts measured 4 times each) for the same gage. If the averages of the two sets are equal then the one with more repetitions will have less STDV right?

Doesn't this mean that I can control the output of my GRR study changing the sample size? What if the first set (4 and 2) failed and the 2nd one (4 and 4) passed the 30% criterion thing, which one should I believe?

Thank you all for your time and comments


M.Bargh

I believe that the whole point of a Gage R&R is to make sure the gage is functioning properly and has the capability to test the item being measured. So you have to have a sampling that will include defects. If you have no defects you cannot statistically prove that your gage is capable. The only variation that you should (ideally) have is the samples. Your operators should be able to be consistent as should the gage.

To put GR&R very simply: the object is to measure a gage's reliabilility in monitoring the process through its measurements.

Please see this paper (http://www.gsilumonics.com/process_online_ordering/pdfs/360.pdf) that explains samples and how the values are statistically derived in a GR&R study.

Why on earth we one needs 10 samples to do a GRR? Isn't having too many parts will introduce a component variability that is due to the process that was used to make the 10 parts. I heard, somewhere, that there are 2 types of GRR studies. One is Gage only RR and the other is Gage/process RR. if you are interested in the 2nd one then yeah sure go and choose 10 parts but if we are only interested in validating the gage, then shouldn't we minimize the # of parts. I would think that ideally one should use only one part or 10 "identical" parts if they are available.

Lastly, lets say that we have 2 sets of reading (say 1- 4 parts measured twice each and 2- 4 Parts measured 4 times each) for the same gage. If the averages of the two sets are equal then the one with more repetitions will have less STDV right?

Doesn't this mean that I can control the output of my GRR study changing the sample size? What if the first set (4 and 2) failed and the 2nd one (4 and 4) passed the 30% criterion thing, which one should I believe?
M.Bargh

The number 10 that is typically used in an R&R study is not mandatory. It originated from the R&R forms used in the automative industry. At that time there were no specialized software packages and 10 simplified the math, like a sample size of 5 used in SPC pre-calculator era.

To answer your question regarding influencing the results by changing the sample size, it depends.

If you are using the P/T Ratio to determine whether a gage is acceptable for inspection to tolerance, the part variation does not change the results, so the sample size just impacts the confidence in the means. If your sample size is too small you may get different results by repeating the study.

If you are using %GRR to determine the acceptability of the gage for process control, the variation of the samples (regardless of sample size) must reflect the within-subgroup variation of the process or your results are not valid. Changing the sample size to deliberately affect your results is game-playing. If the samples (5 or 50) reflect the actual process variation, the results will agree.

Select enough samples so that:

Number of samples x Number of Operators > 15


If this is not possible or practical...

s x o < 15 trials = 3
s x o < 8 trials = 3 to 4
s x o < 5 trials = 4 to 5
s x o < 4 trials = 5 to 6

The number 10 that is typically used in an R&R study is not mandatory... If the samples (5 or 50) reflect the actual process variation, the results will agree.

True the number is not mandatory.
and small numbers of parts can be used (particularly if the test is time consuming and/or expensive)
However, one must be cognizant of sampling error. Standard deviations and averages can ahve very large error when using small number of parts. I typically use aroudn 30 parts.

Beautiful replies. I'm now tempted to ask more.

How do I determine which to use, P/T or %GR&R? My process does have specification window that is about 0.6 um but the process itself is not that precise. Meaning, my process could produce 2.1 um when you want it to produce 1.9. Pretty large variation for a specification width of only 0.6 um.

My gage is pretty accurate. It measures as thin as 10 nm. It is really the non-uniformity of the membranes that I'm measuring that is causing all the variation. The operator-operator variation is not small but still smaller than the process variation. The “6” that I have to multiply my sigma with is what could potentially cause (I have not done the study yet) the gage to fail if I'm to use the P/T test.

The paper that Jennifer cited about the %G&R doesn't mention the 6*sigma factor which makes the acceptance criteria for %G&R significantly different from the P/T. AM I right here?

The %GRR as described by the paper, relies on the ratio of the RR to TV which is the sum of RR and PV. How do I determine PV? Is it determined from the sample size or does it have to do with the tolerance?

Thank you all for your valuable input


M. Bargh

How do I determine which to use, P/T or %GR&R? My process does have specification window that is about 0.6 um but the process itself is not that precise. Meaning, my process could produce 2.1 um when you want it to produce 1.9. Pretty large variation for a specification width of only 0.6 um. Use P/T Ratio if the gage will be used to inspect to a tolerance. Use %GRR if the gage will be use for process control (i.e., SPC).

My gage is pretty accurate. It measures as thin as 10 nm. It is really the non-uniformity of the membranes that I'm measuring that is causing all the variation. The operator-operator variation is not small but still smaller than the process variation. The “6” that I have to multiply my sigma with is what could potentially cause (I have not done the study yet) the gage to fail if I'm to use the P/T test.What you describe is within-part variation and it may influence either Repeatability or Reproducibility. If all operators measure randomly, it will show up in Repeatability. If one operator measures the smallest value and another the largest value, it will show up in Reproducibility. If some are random and some measure the smallest or the largest, it will show up in both. Train all the operators to a standard method, such as the maximum/minimum/mean value of x measures to reduce this effect.

The paper that Jennifer cited about the %G&R doesn't mention the 6*sigma factor which makes the acceptance criteria for %G&R significantly different from the P/T. AM I right here? Both measures use the 6*sigma. This is simply part of the the method. It used to be 5.15*sigma, but was changed several years back. %GRR is this value as a percent of the part variation, while P/T Ratio is this value as a percent of the total tolerance.

The %GRR as described by the paper, relies on the ratio of the RR to TV which is the sum of RR and PV. How do I determine PV? Is it determined from the sample size or does it have to do with the tolerance?

PV is determined by the variation of the parts. The samples that you use, whether 5 or 50, must reflect the actual process variation. Tolerance has no impact. Yes, you can deliberately influence the PV by the number of samples as well as how you collect the samples. However, you should collect the samples in a way that you see the actual process variation.

hi jennifer' can help me on my first job.. because i am on Bussiness Improvement Team.. so what would be my Job details?? thank you'

I just have one quick note on the paper poited by Jennifer.

The paper says that %R&R is given by %R&R = 100[R&R/TV]. If the result of this formual is a percentage, then it should really be

%R&R = 100[(R&R)^2/TV^2] rather than the formual above. This will ensure that

%R&R + %PV = 100

Does anybody have a take on this. Am I right?

M.