GRR results not match with Minitab results

[#1]

Hi, recently I received a GRR results file from a customer who used Minitab software to calculate it. But we cannot get the same result when we tried to use JMP14 software. Please give your suggestions on how to get the same results. Thanks

[#2]

Could you post the results of both analysis.  How different are they? There are many ways to estimate the components of a measurement system analysis.  The control chart method (See Wheeler), ANOVA, and it can depend on whether the study is crossed, nested, or a combination.

[#3]

Hi,

Thanks a lot for your feedback and great help.

I attached the raw data and Minitab results for your reference. Minitab uses Nested model.

[#4]

First, you have what looks like a typo in your data set (row 71).  Second, Operator and Part are crossed and the repeats (run) are nested. Minitab is showing 27 DFs for part and yet you have only 10 parts in the study.  This is incorrect.  The numbering scheme in your data set also appears incorrect.  For the Run column, I believe this is precision repeatability, the numbering should be 1-90 unless you have some hypotheses about a possible systematic pattern to the order of measurement (every measurement is unique).  I believe you can have Minitab do a crossed study. I have attached a .doc with that analysis.

That being said, to be honest, I've never used the JMP Gage R&R (Option under Variability Charts) platform (I'm old school and do the analysis using variability plots, traditional control charts and ANOVA).  The Variance Components (Variance Components option under Variability Charts) appear correct.  The control chart appears correct.

However, when I tried the Gage R&R  option, I got strange and apparently incorrect, results (Perhaps I am doing the analysis wrong??? ). Examining the Variance Components for Gauge R&R, there is no way the Reproducibility is 99.33%.  This appears mis-labeled (I would report this to JMP technical support).

I have attached a journal (and the JMP data table) where I first did the analysis as a nested study (which is incorrect).  Then I repeated the study as a crossed-nested study.  The variance components look correct.  The Gage R&R options look to be wrong?

[#5]

And I'll put in a plug for the correct results. Just because another application produced a result, doesn't mean it's correct or the optimal way to perform the analysis. As @statman asks, without know exactly how the result was created in Minitab, the experimental design and the hypotheses you may be testing as well as what the customer would like to learn or conclude...we can't be of much help.

[#6]

Hi Bartell,

Thanks a lot for your feedback and great help.

I attached the raw data and Minitab results for your reference. Minitab uses Nested model and I tried Nested and Crossed models, but cannot get same results.

[#7 · Accepted solution]

Hi @JKumar,
I ran GR&R on your data set and one of your data points (71) had an extremely high value. With that point excluded the JMP results were closer to the Minitab results, but did not match. I put the data point back in with a decimal point where I thought it should be. The results were closer, but still different.  I then reran the analysis, but this time I went to the Analysis Settings and chose Best analysis (EMS or REML) and these results came up much closer to the Minitab results.  Please see the image below.

Based on these results I would say the difference in results lies somewhere in the difference in the methods/algorithms used in JMP and Minitab and maybe differences in the data used for each analysis.  

Retry your analysis by changing the Analysis Settings under Options in the Measurement System Analysis Platform, and doublecheck that the data matches exactly for both Minitab and JMP.

HTH

Bill

[#8]

Hello,

I wanted to make a quick comment on Bill Worley's response.  I believe that the reason we see differences between JMP and Minitab in Bill's GRR report is because JMP is estimating the variance components using REML and Minitab is using EMS (expected mean squares/ANOVA) .  REML and EMS results will be the same if the data is balanced and there are no negative variance components that need to be zeroed, but that is not the case here.  

I hope this helps to explain the differences a little bit.

Laura

[#9]

Hi Laural,

Thank you for the great help.

Is there any way to select the EMS option manually in the analysis setting?

J Kumar

[#10 · Accepted solution]

Hi,

There is no way to "force" EMS within the Variability platform because REML gives better (less biased) estimates when there are negative variance components that need to be zeroed.  However, it is possible to do EMS in the Fit Model platform in JMP.  If you use the following script with the JMP data table you can get the EMS estimates for this random effects model:

Fit Model(
	Y( :Data ),
	Effects,
	Random Effects( :OP, :Part[:OP] ),
	NoBounds( 0 ),
	Personality( "Standard Least Squares" ),
	Method( "EMS" ),
	Emphasis( "Minimal Report" ),
	Run(
		:Data << {Summary of Fit( 1 ), Analysis of Variance( 1 ),
		Parameter Estimates( 1 ), Lack of Fit( 0 ), Scaled Estimates( 0 ),
		Plot Actual by Predicted( 0 ), Plot Regression( 0 ),
		Plot Residual by Predicted( 0 ), Plot Studentized Residuals( 0 ),
		Plot Effect Leverage( 0 ), Plot Residual by Normal Quantiles( 0 ),
		Box Cox Y Transformation( 0 )}
	)
);

Notice that the OP variance component is negative in the Fit Model EMS report.  It just needs to be set to zero to match Minitab.

I hope this helps.

Laura