Calculating Cpk on Non-Normal Data Distribution

[d.conroy]

Hi All,
So I've been doing validation work for a while now so I'm resionabilly fimilar with CPK and other process indicators. However our ongoing validation has lead me into a problem.

I'm calculating a Cpk on a heat sealing process based on seal peeling strength evaluations. There is a lower limit of 1.5N/15mm strip with no upper limit.

When I look at the results there is a mean on 3.4N with a SD of 0.78, so not good; cpk of only 0.8:

But then I look at a distribution plot of the results and see that the distribution has a large tail on the positive side of the mean.

This tail is giving me a large SD which is causing a low CPK.
This dose not seem reasonable to me; the outliers/tail above the mean is making me fail the lower specification CPK.

What can I do in this situation?

If I calculate the SD only on the lower side of the mean i.e. (summation of: [3.4-{result less than 3.4}] /{n of results less than 3.4})^0.5.
I then get a SD of 0.37 and a CPK of 1.7

Is this reasonable to do? What else could I do?

Any help would be appreciated.

 

[Bev D]

yes, this is a reasonable approach. it meets the intent of capability indexes.

 

[d.conroy]

Thanks for the reply Bev

 

[Darius]

I would...

Function DS_dwn(R As Range) As Single
Dim Median As Single
   Median = Application.WorksheetFunction.Median(R)
   ds1 = Median - Application.WorksheetFunction.Percentile(R, (1 - 0.683) / 2)
   ds2 = (Median - Application.WorksheetFunction.Percentile(R, (1 - 0.955) / 2)) / 2
   ds3 = (Median - Application.WorksheetFunction.Percentile(R, (1 - 0.997) / 2)) / 3
   DS_dwn = (ds1 + ds2 + ds3) / 3
End Function

Maybe strange, but it was handy for me, a non-parametrical way to estimate stdev one-way (up or down, in this case is down) using the percentile of the normal distribution. I developed it to modify Chauvenet criteria for outlier estimation.

 

[Darius]

or if you want it by IQR (interquartile range)

Being DATA_RANGE the data range, the ds from the lower tail:

=(MEDIAN(DATA_RANGE)-QUARTILE(DATA_RANGE,1))/0.67448

demonstration:

By bibliographic references IQR=2*(0.67448)ds

if you take only the lower part, you would need to duplicate the Median-1Q value to simulate that the same that happens to the lower bound happends to the upper to simulate a IQR, let's call it IQR'.

so: IQR' = 2*(0.67448)ds
2*(Median-25Percentile)=2*(0.67448)ds

so ds = (Median-25Percentile)/0.67448

 

[Statistical Steven]

Hi All,
So I've been doing validation work for a while now so I'm resionabilly fimilar with CPK and other process indicators. However our ongoing validation has lead me into a problem.

I'm calculating a Cpk on a heat sealing process based on seal peeling strength evaluations. There is a lower limit of 1.5N/15mm strip with no upper limit.

When I look at the results there is a mean on 3.4N with a SD of 0.78, so not good; cpk of only 0.8:

But then I look at a distribution plot of the results and see that the distribution has a large tail on the positive side of the mean.

This tail is giving me a large SD which is causing a low CPK.
This dose not seem reasonable to me; the outliers/tail above the mean is making me fail the lower specification CPK.

What can I do in this situation?

If I calculate the SD only on the lower side of the mean i.e. (summation of: [3.4-{result less than 3.4}] /{n of results less than 3.4})^0.5.
I then get a SD of 0.37 and a CPK of 1.7

Is this reasonable to do? What else could I do?

Any help would be appreciated.

Just to add to the fray....I am not a believer in penalizing yourself for having a long tail away from the specification. There are several approaches I use:

1. A log transformation to get the data more normal.
2. A method similar to Darius, a non-parametric Cpk metric.
3. Do a 5% winzoring of the data. That is remove the upper and lower 5% of the data. How does the Cpk change.
4. If want to do some heavy lifting, you can use non-parametric tolerance intervals when the sample size is large. I can then predict how many PPM outside the lower specification I can expect.

There are many ways to handle non-normal data for capability, but as Bev said, it's about remaining true to the intention of capability. At this point in time do I have evidence of a capable process.

 

[Bev D]

some very good suggestions on how to handle this situation statistically. As a stats geek I find them really interesting - but only in an academic way. My practical side lets out a large sigh and says "what a lot of work for such a meaningless index".

I would ask a completely different question at this point: why do need to calculate a Cpk index at all?

 

[Statistical Steven]

some very good suggestions on how to handle this situation statistically. As a stats geek I find them really interesting - but only in an academic way. My practical side lets out a large sigh and says "what a lot of work for such a meaningless index".

I would ask a completely different question at this point: why do need to calculate a Cpk index at all?

Bev, let me give you my perspective, though some might not agree. the use of Cpk as an acceptance criteria is an off-shoot of six sigma. Instead of understanding process variation, this method forced you to answer "How many PPM can I expect". As I tell people, you can measure Cpk today, and then again a week later and see different results. Even if you use Ppk is does not change. I asked a coworker to compare a Cpk of 2.0 with the mean was very close to the specification with low variability to a Cpk of 2.0 with the mean centered but more variability, which is better? The answer was they are the same. This is why it's a meaningless index!

Thanks for your valuable input.

 

[Miner]

Bev, let me give you my perspective, though some might not agree. the use of Cpk as an acceptance criteria is an off-shoot of six sigma.

Not an offshoot of Six Sigma. I was seeing this from the automotive industry in the early 80's. Six Sigma didn't gain much attention outside Motorola and the early adopters until the mid 90's.

 

[Bev D]

Miner is correct that Cpk came over from the Japanese in the mid 80s (Sullivan, L. P., "Reducing Variability: A New Approach to Quality", Quality Progress, July 1984)

The automotive industry unnecessarily complicated the original concept by introducing the concept of short and long term variation (Cpk & Ppk)

Steven is correct in that some Six Sigma folks started to corrupt Cpk with a bunch of nonsense about:

• being able to predict very low defect rates based on the tails of the Normal distribution
• espousing that the difference between the long term and short term capability represented the entitlement of the process; preaching that assignable causes are easy to eliminate and that common causes are not.
• and then came the 1.5 sigma shift abomination
• and then insisting that everyone had to have Cpk values for every characteristic or they weren't really 'doing' six sigma...

===========================
if you are interested in the defect rate (for cost and yield impacts to delivery, or fro understanding the inspection needs) then you are far better off counting the number of defects that actually occur. (Pyzdek, Thomas, "Why Normal Distributions Aren't [All That Normal]", Quality Engineering 1995, 7(4), pp. 769-777)

If you are interested in the overall variation of the process in relation to the specifications, you are better off plotting the process results in a multi-vari chart and comparing to the specification limits. (it is insensible to reduce variation to a single number)

If you are interested in the future stability of the process you are better off plotting the process in a multi-vari and/or control chart and understanding the science then applying the appropriate controls to maintain stability and detect changes quickly. (statistics doesn't obviate science)

Cpk (and all of its illegitimate children) simply can't do any of the above for you...