Capability analysis

[#1]

Hi have continuous data (processing time)

The Continuous Fit suggests the "Normal 2 Mixture Distribution"

Can I make a capability analysis with this distribution?

Thanks

[#2]

bump

[#3]

Yes, easily.  In the capability analysis dialog within Distribution, you'll be able to choose a distribution for the data.  Normal 2 Mixture is an option.

[#4]

Cameron

Are you using JMP pro, or maybe it is a new feature in JMP 14? I'm using JMP13.2.1 standard and I don't think I have that feature.

My two cents on the topic is that if I care enough about a given parameter, I would make sure I understand why the distribution is multi modal, understand how close the modes are to one another, whether they're stable, and decide whether it needs to be or can be "fixed." If not then I'd apply some robust stats to it. For example either JMP's robust mean / standard deviation to get a robust cpk, or use a function of median and iqr to do the same.

-Mike

[#5]

I'm using base JMP version 13.2.  You can also get the capability analysis for any of those fitted distributions by clicking the red-arrow icon for the fit and selecting "Spec Limits".  Setting those will generate a capability analysis.

[#6]

Thanks, sorry about that, I hadn't scrolled over far enough. Spec limits were there so I had default cpk, but I didn't see the normal 2's cpk.

[#7]

If the normal mixture model is a valid description of the data, then it implies that there is more than one normal population. What does it mean to ask if they are stable and capable as in the case of one population? That they are individually stable and capable? That the mixture of the two is stable and capable?

Why is there more than one population? Can they be identified?

[#8]

With semiconductor test data we need to consider how the product is performing vs how the test equipment says the product is performing. For large scale testing, there would be many testers used, sometimes different versions, resulting in fairly fixed offsets (multi-modal) for a handful of parameters. If such a parameter specification is critical we can address it separately, or just verify we understand why and set the limits accordingly. I use gauge R&R to quantify this.

"Stable" may not have been the correct term, but as an example of unstable: There might be some problem with 1 piece of test equipment for used for one lot in the factory, where a subset of the data shows up as a new mode. I would not want to include this aberrant data if I was looking to set cpk based limits.

[#9]

"set cpk based limits?"

[#10]

For a given continuous parametric parameter, we'd generally set limits, or pass / fail criteria, based on a target cpk. For example if our target cpk was 2, I would set limits at mean +/- 6 sigma. Or a robust version of the same.

For some types of parameters, the distribution is such that it doesn't make sense to do this, usually cases of test measurement limitation or some idiosyncrasy of the product. Or if it is a resistance measurement, and the statistically set lower limit is negative, I adjust that to a positive value as electrical resistance cannot be negative. In these cases it is simple enough to eyeball some limits, though I generally try to apply robust limits first and if they don't make sense (too wide, too narrow, etc), I'll just manually implement a sensible limit for the parameter. Criticality of the parameter itself is taken into account.