Hi folks! First time poster here.

One of our suppliers recently FAXed over some parallelism data (using DataMyte/Quantum SPC). I manually copied the raw data over into MINITAB (our software package) and found (not surprisingly) the data is not normal (A^2 = 6.95, N=0.000). I was taught that if the data was not normal I should not do any further statistical analysis.

I have no reason to believe their Quantum data is invalid, but I cannot confirm their Ppk values using MINITAB. What to do?

Thanks in advance for your thoughts.
Kevin

Hi folks! First time poster here.

One of our suppliers recently FAXed over some parallelism data (using DataMyte/Quantum SPC). I manually copied the raw data over into MINITAB (our software package) and found (not surprisingly) the data is not normal (A^2 = 6.95, N=0.000). I was taught that if the data was not normal I should not do any further statistical analysis.

I have no reason to believe their Quantum data is invalid, but I cannot confirm their Ppk values using MINITAB. What to do?

Thanks in advance for your thoughts.
Kevin

The old "If the data are not normal..." is not valid. However, you do have to be careful to use the correct statistical techniques.

Minitab has an option to analyze non-normal data either by distribution fitting or by data transforms (to normalize the dats). Note: it is perfectly "Normal" for parallelism data to be non-normal. You are dealing with what is called a "Boundary" condition (i.e., you cannot go below zero). Minitab can hadle this also. Enter 0 as the lower spec and check the Boundary box.

Man, first you can try to identify which distribution fits best using Quality Tools/Individual Distribution Identification.

After that you can use the Non-Normal Capability analysis selecting the distribution you've found in the previous step.

Also you can use a Box Cox transformation in Normal Capability and using 0 as the lower spec and checking Boundary.

Good luck.:bigwave:

Regarding 'boundary' you both are correct. I checked the box and re-ran MINITAB and I now agree with my supplier's Ppk. Thanks guys.

Regarding 'normality' I am still, well, unsure. In this instance I understand parallelism is a unilateral tolerance (can't be less than zero). So...

1. Should I just 'accept' the fact that the data is non-normal and predict Ppk?

2. Are there other MINTAB functions that I can apply to a given set of 'non-normal' data?

My first inclination when viewing the data was, 'ah, they're cherry-picking their points.' All opinions, URL's, enlightenment, or gifts of money would be appreciated.

1. Should I just 'accept' the fact that the data is non-normal and predict Ppk?

2. Are there other MINTAB functions that I can apply to a given set of 'non-normal' data?



VMax, you can't apply regular capability analysis to non-normal data. You will have to transform it with Box Cox or identify which distribution matches best to then use the Quality Tools/Non-Normal Capability Analysis of MINITAB.

If you use regular capability analysis to non-normal data you'll get the wrong information. :yes:

I will recommend www.isixsigma.com forums for these topics. Do a search or ask there and you'll get a ton of replies in minutes.

Regards.

Many software including MINITAB obtain the ppk no matter if "is normal" or not.

The first step to compare results, check if MINITAB obtain the same value without any transformation (I am pretty sure that the results will be equal), and if not is equal, try first to see if changing options in the configuration screens make them equal (A friend of my obtained two different values in 2 differents reports of MINITAB and checking the options changed both to the same value).

Do not use a lower spec of 0 if is not the case, the indicators like ppk take in account that the target is centered on the specs.

Agree that a transformation (and I really like Box Cox) could transform your distribution an the value obtained will represent with more reality the data but as I think "If you are going to jump into a pool is better to try to see the bottom of the pool before jumping than not doing it because you have no rule to measure it.". An Indicator by it self is better than none.:mg:

note: if you are going to transform the data, you will have to transform the specs also.:caution:

1. Many software including MINITAB obtain the ppk no matter if "is normal" or not.

2. Do not use a lower spec of 0 if is not the case, the indicators like ppk take in account that the target is centered on the specs.

3. note: if you are going to transform the data, you will have to transform the specs also.:caution:

Darius,

I will comment what you've just said.

1. Yes, MINITAB will output a Ppk result no matter what distribution you input but YOU WILL GET THE WRONG INFO! :mad:

2. He said that he is using a set of data from a parallelism test. That will have a boundary of 0.

3. The specs will be automatically transformed by MINITAB. Maybe you meant that the result will show transformed specs, but you will be looking at the Ppk values only.

Regards. :agree1:

A few additional thoughts:

1. Make sure you are using the latest version of MINITAB, which is MINITAB 14.20. Release 14 has a much much more powerful set of non-normal tools than earlier versions, such as release 13.

2. The capability tools do handle one-sided distributions. Just leave the non-applicable spec limit blank. Only check the boundary checkbox if it is physically impossible for a value to fall below or above the boundary. The boundary checkbox is used when the PPM values are calculated, excluding the distribution on the other side of the boundary from PPM calculations.

3. As mentioned several times before, make sure you use the Individual Distribution Identification tool to find the distribution that fits your data best. If that doesn't work well, try the Johnson Transformation option inside the Nonnormal Capability tools.

4. Don't forget that MINITAB users can always call Minitab Technical Support. Except for the cost of the call, it is free to all users. For the number go to Help > About MINITAB and find the Contact Us button. They are very helpful.

I called Mintab this AM. They sent my company a v14 upgrade CD and it has apparently disappeared into the IS department. I'm trying to sort that out right now.

v13 doesn't have the features you folks are describing (Non-Normal Capability Analysis, etc.), so for the short term I'll d'load a 30 day demo of v14. IS should respond within a month and get the upgrade installed on the network.

All of this work for 0.2mm...

-V-

As other posts, the start point looks one way

I have no reason to believe their Quantum data is invalid, but I cannot confirm their Ppk values using MINITAB. What to do?


And looks now about the best way to calculate ppk....

Do you have any reason to believe that Quantum is making ppk using non normal capability index?

Can you post your data ( a sample ), and the ppk resulting from Quantum?, maybe somebody can use "the non normal capability analisys" and show if Quantum is using such analisys or not.

OK. I downloaded/installed v14. Input the data (http://www.geocities.com/junkvmax/raw_data.xls). Ran Quality Tools/Individual Distribution Identification. The charts represent 'best fit.' Extremely low or nonexistent P values. AD ranged from 16.4 to 2.8.

Q: For non-normal distributions (specifically parallelism), are there threshold numbers for P and AD? Why/why not. I have been taught for normal distributions, AD<0.752 and P>0.05. Is there a similar 'rule-of-thumb' for non-normal data?

In this example, 3-Parameter Gamma contains all the points and has the 2nd lowest AD value. 3-Parameter Lognormal has the lowest AD, but some of the data points are outside the lines.

Q: Should I base the next step (Capability Analysis/Nonnormal) on the graph or the values of AD and P? I've tried it both ways (3-P Gamma and 3-P Lognormal). Similar Ppks.

I included that data - link above. 230 data points.

Extremely low or nonexistent P values.

VMax. I tried your data. If P-Value is less than alpha then none of the proposed distributions fit your data.

What I would do is go with a Box Cox transformation in Normal Capability and get your info.

What is your upper limit?

Pablo :bigwave:

If P-Value is less than alpha then none of the proposed distributions fit your data.

OK. Thanks for the tip (and taking time to review my data). Tried Box Cox Transformation and got this error:

* ERROR * All data must be positive when using the Box-Cox transformation.

I changed my data (replaced the 0's with 0.0001) and finally got the Box Cox to spit out a graph. Is this permissable?! The USL is .009.

OK. Thanks for the tip (and taking time to review my data). Tried Box Cox Transformation and got this error:

* ERROR * All data must be positive when using the Box-Cox transformation.

I changed my data (replaced the 0's with 0.0001) and finally got the Box Cox to spit out a graph. Is this permissable?! The USL is .009.

No. If you have to change numbers from zero to make the distribution work, you are breaking a primary assumption of the distribution, which proves that the distibution does not apply.

I used Minitab to run an Individual Distribution Identification. Four distributions offer potential answers (3-parameter lognormal, 3-parameter Weibull, Largest Extreme Value and 3-parameter Gamma) by the Anderson-Darling statistic. The low p-value rules out the 3-parameter Weibull; the probability plot rules out the Largest Extreme Value; and, the probability plot is very questionable for the 3-parameter Gamma.

The best fit based on A-D, p-value and the probability plot appears to be the 3-parameter Lognormal distribution.

You can then use a non-normal capability analysis using the 3-parameter Lognormal distribution with 0 as a lower Boundary and .009 as an upper spec. This gives a Ppk = 1.09 and an expected 863 PPM overall performance.

I changed my data (replaced the 0's with 0.0001) and finally got the Box Cox to spit out a graph. Is this permissable?! The USL is .009.

VMax, I don't find it a big issue. You are simply using a small number that is not zero but very close to zero to allow the Box Cox transformation to run.

Compare the results with what Miner said in the previous post.

Pablo :agree1:

I don't have any stat pack but I checked with X^0.5 and the standard deviation looks almost the same both ways (up and down), the ppk that way is 1.2. The box-Cox obtain the exponent for the X, so it looks like 0.5 to me (in your set of data).