GRR (Gage R&R) with Different Scales

[findthebird]

Hello!

I would like to make a GR&R (with STATISTICA) for a device which measure the power of optical lenses. I have data of 15 samples, with 5 trials, 3 operators. BUT the samples' range is from 0 Diopter to 30 diopter. Since this, the part-to-part variation doesn't make sense.
I have nominal values for these lenses. My thought was to take away the nominal value, but i'm not sure. (And there is also an other problem: the nominal values made from the measured. The gauge's resolution is 0.01 Dpt and from that data we round the measured values to nominal classes with 0.5 range.)


Any ideas?

 

[Stijloor]

A Quick Bump!

Can someone help?

Thank you very much!!

 

[Ninja]

Howdy findthebird,
Welcome to the Cove!

What are you trying to accomplish with the GRR?
Your own peace of mind? Compliance with a standard? Meet customer requirement?

Are any two or three (or more) of the lenses tested made to meet the same nominal diopter?

What are the acceptability limits (tolerance)?

Can you post the data?

Nuff of the questions...I will often run two different parts in a 3x3x10 GRR for testing the gage, then break the data set into (2) 3x3x5 sets to look deeper into the part variation itself. The part variation is HUGE when I have dissimilar parts in the same set...but that's OK since I know why the variation is so big. The subsets are looked at for info on the variation among similars.
But it all comes down to what you are trying to get out of doing the GRR...

 

[findthebird]

Thanks for the reply.

After measuring the lens, we categorize it. As I mentioned, the resolution is 0.01 Dpt, but the classes are with 0.5 Dpt. So there are no classical acceptability limits, because after 1 measure, we decide if the lens is in the nominal class, or we have to shift it to an other due to some difference in the raw material, or some manufacturing cause. So basically there is no reject because of the result, for that there are som other criterias.

We have 3 devices. My goal to prove: the devices measure the same results for the same lenses, with the same error. And the operator (and the skills or the temper of his or her) has a big influence on the result.
I have 3 op., 3 device, 5 trials and 15 samples. The samples, as you can see in the attached file, are quite different, so i think i cannot make classes for them.

P.S.: The data is ready for STATISTICA, so sorry for the format.

What is your oppinion?
Thanks in advance!

 

[Ninja]

Wow...thats a lot of data. We usually see people asking "Do we really need to measure that many times?"
You have the opposite: "How many ways can I slice up the huge data set?"

I ran a quick subset 3x3x10 using the first three operators, first device, first ten measurements, Tolerance of 0.5 (from your posts above) and got:
EV% = 9.2 equipment is near the edge (gage 1)
AV% = 19.3 appraisers using this device vary a lot from user to user
%R&R = 21.4. Below 10 is most desireable.
PV% = 1362.7 -- a meaningless number since you are using known different parts.

You can slice your data set into multiple checks to test various hypotheses:
- Are all the devices equivalent? (slice out a 3x3x10 using 3 people, 3 devices, 10 measures and look at EV%)
- Are all the operators "equal" on a given device? (other than Device 1)
- Are operators more equivalent on one device than another? (run 3x3x10 on each gage and compare results)
- I didn't look at same-part stats, but among the 15 parts, are two or three of them near identical? If so, especially if you have more than one "group of similars", is one device more repeatable for one range and another device more repeatable for another range?

In the end, you may find a segment by segment statistical analysis more useful than a GRR...unless you specifically need the GRR for some other reason.

HTH

 

[findthebird]

Very big thanx!

I will do, as you recommended, but before that, while I read your post, 1 more thing came up to my mind: Operator 1, 2 and 3 are the same with the 3 device. Is there any sense to look from an other perspect of view? Let's say, the devices are the operators and they are "using" the handler. Can I get something different from it? (Now, when i'm translating my thoughts, i think, it would do the same, but I'm asking it anyway...)

 

[Ninja]

I agree, there are a few different ways to mine your data to look at the data set from different angles.

- Do the three devices give the same result for one user?
- Do the users get the same result from one device? (answer is "NO" for device 1, what about the other devices?)
- Do the answers to the above two questions change with different ranges of diopter?

Look at all of the angles related to what you are trying to achieve and see if you can draw a pattern or a trend out of the results that helps to direct your path forward.

 

[findthebird]

Sorry, i was occupied with other subject, but thank for your answer!

Now i am confused. I didn't get the same data. For me, with the G1, first 10 measurements with 3 operators, 5 trials, 0.5 tolerance applying Range method:
EV%: 28 %
AV%: 44 %
%R&R: 52 %
PV%: 3507 %

And with ANOVA:
EV%: 32 %
AV%: 40 %
OxP%: 45 %
%R&R: 68 %
PV%: 3210 %

These are the % of the tolerance.
(I use STATISTICA)
And there are big differences between yours and mine. Did i misinterpreted sg?

 

[findthebird]

OK, i'm thinking constantly

If I round the data, I got from the machine, for making a decision for the nominal value, shouldn't I round the measurement data also for the MSA?
So maybe the task is a little bit easier, if i round to quarters, because the nominal values are in half step classification? And the tolerance will be +-0.25?

Thanks in advence!

 

[findthebird]

I think I get it. Not specific GR&R what I need in this case, but 3-W ANOVA, wher I should decide, if the gages and the operators are giving the same results.

Am I right?

Sorry, i just started...