Search the Elsmar Cove!
**Search ALL of** with DuckDuckGo including content not in the forum - Search results with No ads.
Ok. So you know, the "classic" study is 10 parts with 3 operators measuring each part 3 times. A total of 90 data points.

This is a good model. Another thing to add - you WANT some variation in your parts. Even better, if a couple of them are out of specification. Remember that you are NOT testing your process, you are testing your GAGE. You know more if you see that it calls a bad part bad as well as a good part good.

Think of it like this: you have to make a quarter inch hole +/- 1/8 inch. And you make a plug gage near those limits. If you decide to test this plug gage and feed it parts that are a quarter +/- a sixteenth, you aren't going to learn anything because you won't see it reject anything.


Starting to get Involved
My next step at the moment is to do the following:
Accuracy Evaluation
  1. Objective: To determine if the measurement tool is accurate and if it needs calibration or adjustment.
  2. Having a valid calibration report (within the last year) will suffice for the accuracy test. If not available, continue with the following steps.
  3. Data collection
    1. Collect 16 measurements on a single standard
Eliminate sources of variation
Same operator, etc.
Take measurements during short time period
Remove standard from tool between measurements (dynamic repeatability)
  1. Analysis
    1. Examine distribution graphically (plot on histogram)
Sanity check (look for outliers)
The cause of any outlier will need to be identified
After sources of outliers are identified and eliminated, re-run accuracy test
    1. Analysis
Calculate 95% confidence interval
Determine statistical significance
If the standard value is inside the confidence interval, then there is not a statistically significant bias.
If the standard value is outside the confidence interval, then there is a statistically significant bias.
Determine technical significance
If the entire confidence interval is contained in the interval defined by the standard value +/- the technically significant delta, then there is not a technically significant bias.
If either limit of the confidence interval is outside of the standard value +/- the technically significant delta, then there is a technically significant delta.
    1. Illustration: The “solid dot” is the mean value and the “line and bars” are the 95% confidence interval

Any suggestions? Sorry I am brand new at this and have to learn on the fly


Bev D

Heretical Statistician
Staff member
Super Moderator
What are you trying to do? This last post indicates that you are concerned about the calibration of a measurement device. Your original post asked about “Guage R&R”. These are very different things.

There are some really good resources about how to perform Guage R&R and calibration. see Miner’s blogs. See my resources at quality forum online in the resources tab, “Practical Quality Engineering Resources section. I have found that the best way to learn is to read them first then try to design a study.

I will also add that mini tab is not a quality tool. It is a software program that allows you to perform statistical calculations quickly. YOU have to know what the appropriate study design is and what is the most appropriate analysis. Minitab won’t do that for you.


Starting to get Involved
Sorry still brand new at statistics. I am trying to do a gage R&R. I am hoping to get 2 additional parts to do this. In the meantime, I have been asked to do a calibration with just the one part. I saw that Excel has a data analysis function. What software do you recommend? I am going to read through the blogs and your resources in the meantime. Thank you for your feedback.
Regarding the calibration: You don't calibrate with a part, you calibrate with a master. If you're making dowel pins and trying to calibrate your micrometer, you don't go get a quarter inch pin from production and measure it looking for 0.2500 in your micrometer. You go get a gage block that is known in size (several, in fact, throughout the range of the micrometer). In truth, you should have a gage block that is traceable to NIST standards. But depending on what your making, you may have some wiggle room. What I mean is, if you're making toothpicks, nobody is really going to care that much. But if you're making bearings, they will.

One more comment about calibrating with a part - IF you are talking about some crazy machine that is designed to only measure a part, what you typically do is make a few "master" parts. These are usually called a go and no-go master and painted green and red and locked in a cabinet somewhere. These are known parts. The purpose of these is - if you bump into your machine with a fork truck and need to know if it still "works" you can pull these out and measure them and if you get the known master results, you are good. People typically run their masters at least once a day if not once a shift (depends on how often the job runs). But here's a key point - that's not really calibration. Your automatic machine is made up of pickups and transducers. THOSE things were (hopefully) calibrated at the factory.

Calibration is all about comparing a gage to known objects to make sure it is giving you a good number.

Regarding Gage R&R: You are stating that you hope to get 2 additional parts. You are correct in that direction, but 2 more is not enough parts. The goal is to have at LEAST 30 or so "data points" for the statistics to work. An industry standard Gage R&R is 3 operators, 3 trials, 10 parts. This winds up being 90 measurements. Why not 10 operators, 3 trials, and 3 parts? Still 90 measurements AND you'll get an R&R out of Minitab if you plug in a study like this....

The reason is you are are comparing noise: noise between the parts, noise between the operators, noise of EACH operator repeating a measurement, noise of the gage, and the noise of the operator/gage interaction. Without going into deep statistics, put your logic hat on .... noisiest components to this are the gage/operator components and NOT the parts. (The parts a physical object, it's not changing measurement to measurement). What this means is, the contribution of the noise from the parts is the hardest to detect, so you need to measure MORE parts. So a 3 operator / 10 part study has a better chance of describing the gage system than a 10 operator / 3 part study.

Now lets get down to reality - you have two things ahead of you. One is increasing your knowledge of these topics. That takes some time. The second is your impending study with bosses (and maybe customers) looking at you, which we need to solve NOW. I don't understand why you can't get more parts, but I don't know what you're making. If we go with constraint that parts are difficult to get - you CAN attempt your study and MAYBE your customer will get off your back. But be warned - if I was your customer and you pulled up some study with only 3 parts in it, you'd be doing the study again in front of me and you would have my full attention. If you have to go down this route, my practical advice would be only show the results page, not the tabular data and MAYBE it will squeak by. You're risk here is ... if your customer digs, he's going to think you are trying to pull a fast one (I certainly would) and you may suddenly find yourself doing Gage R&Rs for his entire product line you are making.

Yes. I am basically telling you your best chance to "trick" your customer. Because that's advice in your current situation. (Very high risk advice). If it works, you need to still go and do a proper study to understand how your gage works. I have burned suppliers very badly when they do a poor study because lack of knowledge of how your gage performs guarantees you will send me a bad part or parts. I don't know your management climate regarding this type of stuff, many managers don't see this as a value add. But if you take my "trick" advice - do so with the knowledge that if you get caught by a customer with shady stuff going on, that results in a lot of future pain for you and your team. And let's be clear - EVEN IF you heart is in the right place, but you do something wrong due to lack of knowledge, the customer is going to interpret that as "shady." We are a paranoid bunch.

The "above board" course of action is to GET MORE PARTS. It is 10 parts. It is OK to include setup parts, it is OK to include bad parts. It is even OK to distress good parts and make them bad. You are testing your gage's ability to detect bad from good.


Starting to get Involved
Thank you for taking the time to explain all of this to me. What happened is the following: My former boss initially asked for 10 pcs from the customer and offered to supply the 10 parts ourselves. The customer wanted us to use one of their parts for the study but stated that they only had 3 obsolete parts for the study. Their engineer agreed to 3 parts and then my former boss agreed to it. Then sometime after the agreement was made, my boss left the company, I got promoted to his position and inherited whatever it was he was working on etc. So here I am. We do have a ring gauge that we can use as a standard. The customer for now just wants to see some data from our CA location which is where the data is from. I did do a Type 1 Gage Study on the part as Miner suggested and a 95% confidence interval on one feature of the part just to see what the results will be . Now I am looking to interpret the graph results to see if the data if I am looking at good results or not. Here is one dimension's results.


Starting to get Involved

I used the mean as a reference even though I don't know what the true value is but this graphs looks better. We will use the ring gauge in order to calibrate. I also got a gist of what the values mean on the Type 1 study from Minitab. Thanks for all your help. I have a lot of research to start on and words of wisdom to help guide me.
I absolutely refuse to use Minitab. So I'm not the best on how you got these output screens.

What I can say is - I'm not seeing the standard Gage R&R output table or a value called Number of Distinct Categories. Also, there's no linearity plot. I know Minitab can spit these things out. If on your last post %Var is % Study Variation, this is telling you your gage can tell the difference between your parts. You should also spit out the % Tol values. That's what you need, which will tell you if your gage is adequate for the tolerances.

Personally, I wouldn't trust Cg/Cgk as a metric at all. I'm not familiar with them, and if I had to guess, I'd say some kid vying for a masters in math or statistics attempted to shoe horn a gage study into a capability study to unify the metrics. Then got a job at Minitab. :) I can't speak to those at all.

And if you are saying you have a feel for the outputs, be damn careful about the PValue. Lots of people say they understand that one, few of them actually do.

We need us a Minitab user to help you more. I'm an Excel guy. I like to type in my own formulas.

Edited to add: Also, there is no indication as to the arrangement of the study. I see an n=33 in there. Is that 1 part measured 33 times? Is that 11 parts measured 3 times? Or one time but in 3 locations? That's key information regarding interpreting the results.

In all honesty - were I the customer, this would look suspicious to me because the full story isn't represented clearly.

Top Bottom