Compression Spring Force measurement correlation with Supplier

[Road Warrior]

Hello all,
I am trying to vet our internal spring force tester, as well as our supplier's unit, and determine if we have correlation. Does anyone have an opinion on the protocol to go about doing this? I'm getting internal disagreement, and would like your opinions. Any calculation tools you could share would be most appreciated as well. I work for a major Tier 1 Japanese automotive supplier, so AIAG tools use is virtually non-existant.
Thank you in advance.
RW

 

[Proud Liberal]

Have you established a set height for the measurement? You'll need the customer to agree to a fixed height if you hope to correlate.

 

[Road Warrior]

Yes, the customer prints give various loads at corresponding heights for each part number.

 

[allan-M]

I've read that some people use a Bland-Altman plot in these circumstances. There are differences between correlation and agreement, and they describe this in their papers. Have not used it much, but I am always on the lookout

 

[Miner]

Read my blog post on Comparison of Two Gages and post #2 of this thread on correlation between a new and existing gage. These should help you.

 

[Road Warrior]

Awesome treasure trove of information Miner - thank you! Now I've got some reading to do.

 

[Bev D]

Miner references the appropriate methods. Youden plots (Dorian Shainin named this the iso-plot) and Bland-altman are based on the same statistics; they just display the data in different formats. the bland-altman does give a direct calculation of the bias between two gages. The job aid Miner references is here: "MSA tools for Elsmar" The spreadsheet includes references for all methods.

remember to establish repeatability of each gage before comparing gages.

What most people gloss over is that standard correlation / regression can't be used for measurement analysis. in standard regression the 'cause' or input factor (x-axis) is taken to be a true value and the variation is all in the output Y value. so the standard deviation of interest is the variation only in the Y value (along the axis). In measurement comparisons both values are variable, so the standard deviation of interest (the measurement error) is on a vector that is perpendicular to a 45 degree line drawn thru the origin. (if there is no measurement error, the points will fall on the 45 degree 1:1 line)

 

[Miner]

What most people gloss over is that standard correlation / regression can't be used for measurement analysis. in standard regression the 'cause' or input factor (x-axis) is taken to be a true value and the variation is all in the output Y value. so the standard deviation of interest is the variation only in the Y value (along the axis). In measurement comparisons both values are variable, so the standard deviation of interest (the measurement error) is on a vector that is perpendicular to a 45 degree line drawn thru the origin. (if there is no measurement error, the points will fall on the 45 degree 1:1 line)

You can use Deming regression, (in Minitab, its called Orthogonal Regression) which takes the variation in the input factor into account, but you still have a problem with interpretation.

In the ideal situation, the equation should be Y = X. That is, the intercept should be zero and the slope should be 1. In reality, you get an intercept that is non-zero and a slope not equal to 1. How do you determine whether this is acceptable or not?

Bottom line, the methods discussed above are better analytical and decision tools than regression.

Yes, this is necromancy of an old thread, but someone thanked my post and made me re-read the thread.