Sample preparation for Gage R&R study - Heterogeneous data or Homogeneous data

[Anatta]

Hi all,

Need some advice from all the MSA guru here...

when it comes to sample preparation for GR&R study, if we collect a heterogeneous data (ie, collect diameters of different models of pen), we will have a significant PV and ndc will surely be good, lets call these method A. Just wanted to understand more whether this is "statistically right"?

the argument was, if you study repeatability, you are looking at; shooting at the target with the same spot as much as possible. when the targeted spot data is collected, you will have a mean value, lets call it 'M'.

by using X-bar R method, the mean should reflect these mean that we are studying, which is 'M'. Now let's name this method B.

If we go back to method A, a heterogenous data with huge PV, will have many sub-mean value for each data. Thus the grand mean for a heterogeous data does not represent the situation compared to method B.

A lot of time, when we use homogeneous data, we can't get very good ndc (which should be more than 5). The challange is to prepare a sample of 10 which is homogeous, yet be different enough to sense by the equipment.

we can cheat by preparing a sample with huge range from a heterogenous data to get good ndc value, but is it defeating the purpose of GR&R study?

next is the argument on AIAG's requirements of meeting ndc>5. In automobile industry, especially when it comes to mechanical measurements (ie the gap between a car door and the chassis ) it has very big tolerence unit. (for case of the car door, we are talking about +/- 1cm!!!), thus is the target value is 2cm, you'll get min value as 1cm and max value to 3cm. When measured, data will be like 1.1, 1.4, 1.3,1.8,1.9, 2.0, 2.4, 2.6, 2.3, 2.8,2.7, (here it looks like a heterogeous data in a homogenoeus data situation, where seems like we have various mean of 1.2, 1.8, 2.5 & 2.75)but in semiconductor industry, we are talking about microns meter...data will be like 0.00012, 0.00011, 0.00011, 0.00013, 0.00011...and so on...the challenge is hard to prepare sample like the ones in automobile industry to get a good ndc...

so, is it right to perform GR&R with heterogenous data or homogenoues data (to be statistically correct, as in X-bar R method, we are stuying only 1 mean).

please enlighten

 

[Marc]

Any takers on this MSA Gage R&R Data post?

 

[Jim Wynne]

It sounds like you're trying to manipulate samples in order to fit AIAG guidelines that are not necessarily appropriate for what you're trying to do. Why would you want to mix up a sample with different parts? It's possible, you know, if you're actually talking about pens or something similar, which isn't clear from your post, to use one particular model/diameter in a study, and use that study as a surrogate for all of the other models.

 

[Statistical Steven]

I cannot speak to the AIAG document, but I can speak to statistically better methods. It is always desirable to have both heterogenuous and homeogenous samples withing a gage study. The best design has samples that span the range of the process.

If you took different diameters and measure them each 5 times, you can get an estimate of within part and between part variance.

The model I use is total variance = part + operator + part*operator + within part.

Hope it helps some.

 

[Jim Wynne]

I cannot speak to the AIAG document, but I can speak to statistically better methods. It is always desirable to have both heterogenuous and homeogenous samples withing a gage study. The best design has samples that span the range of the process.

If you took different diameters and measure them each 5 times, you can get an estimate of within part and between part variance.

The model I use is total variance = part + operator + part*operator + within part.

Hope it helps some.

Steven, I think when the OP is referring to "heterogeneous" she means different part numbers--in other words, different nominal sizes and perhaps even different tolerances.

 

[Anatta]

Jim,

you are right. our organization has engineer using a range of sample that covers the process, that of different models, in these situation it's a heterogeneous data, where you have more than one mean value.

for the case i mentioned above, then the mean value should be which of the many "means" we should using when plotting the X-bar chart.

i am having a hard time convincing them that an GR&R study should use a homogeneous data. The set of sample to provide the measurement should have only one target mean value. so, when every measurement is made and whether it fall in the range of the mean value, then you can determined the precision. whereas, if you have a heterogeneous sample, which are the mean you are refering to, when we try to study the precision (a.k.a repeating the measurement on the same spot on the bulls eye!!!a hetero data will have many target, take an average of the many target, how are we going to study the repeatability...which is repeat at the average value of an extreme range of data is wrong)

need more comments and advice here...

 

[Anatta]

Re: Sample preparation for Gage R&R study - Heterogenous data or Homogenoues data

to add more...

the ndc value, is to discriminate the measurement system, or to judge the sensitivity.

and, the if the rule of thumb on 1/10 of resolution is use.

we should perform a GR&R with homogenoeus data, to see...whether our measurement system, is sensitive enough to distinct or discriminate the "1/10" changes in the samples that we have prepared. (the samples are prepared homogenouesly and yet still within it's tolerence range with each samples or >5 of it are at least 1/10 of a resolution different from each other to test whether our measurement system can "catch" the difference).

my understanding may be wrong , please advice...:truce:

 

[Miner]

Mixing parts of different nominal sizes is an absolute no-no when doing a gage R&R study. The intent is to obtain sample parts that reflect the process variation for a given nominal size to compare with the R&R of the gage.

Mixing the nominal sizes is no more valid than performing a capability study that includes different nominal sizes. The only difference is that no one would do this because it makes the results worse. Whereas people want to mix nominal sizes for a gage study because it makes the results better (though falsely).

 

[Statistical Steven]

Re: Sample preparation for Gage R&R study - Heterogenous data or Homogenoues data

Mixing parts of different nominal sizes is an absolute no-no when doing a gage R&R study. The intent is to obtain sample parts that reflect the process variation for a given nominal size to compare with the R&R of the gage.

Mixing the nominal sizes is no more valid than performing a capability study that includes different nominal sizes. The only difference is that no one would do this because it makes the results worse. Whereas people want to mix nominal sizes for a gage study because it makes the results better (though falsely).

I think this is an issue of the AIAG guidelines versus statistical analysis. I have done gage studies using different nominal sized parts to get the full range of the measurement equipment. Using variance component analysis, you can quantify the part to part variability (which is HUGE and not important) from the gage error associated to the operator, gage and the important operator by part interaction. These standards give you metrics for a gage study, but if all you want to know is if the gage is capable for the process, then you look at the gage variance sources versus the nominal values of the different parts to see if it is capable for any given part.

 

[Miner]

You are correct in that if you specifically design the study and analyze it as you suggest that you can get useful and correct results from it.

My answer was from the perspective that a very large percentage of people use the AIAG cookbook approach for the analysis. In this situation, a heterogeneous sample will inflate the process variation as determined by the AIAG method and provide a %GRR that is much lower than it actually is.

If the experiment is designed for a specific analysis, and the ANOVA is performed to separate the impact of the different nominal dimensions, you can obtain accurate results.