# Non-normal Process Capability-Help!

Six Sigma – iSixSigma › Forums › Old Forums › General › Non-normal Process Capability-Help!

This topic contains 8 replies, has 8 voices, and was last updated by Salomon 14 years, 8 months ago.

- AuthorPosts
- October 21, 2004 at 4:10 pm #37290
I have a data set that is not normal. I ran a probability test on it and the curve that best fits is the Logistic curve at AD* .76 with Normal next at AD* 1.51. How do I get a process capability for this data set. I see an option for normal, weibull, between/within, poisson, and binomial. Do I just assume normality? Help! Thanks

0October 21, 2004 at 4:19 pm #109508Rob,

I would start by first trying to understand why your data isn’t normal. Do you have some boundry condition, is it time related??

I wouldn’t just blindly transform your data in order to calculate some capability metric. Understand why your data looks the way it does.0October 21, 2004 at 4:41 pm #109510It has a boundary limit of 0. I am also not sure of what the Goodness to Fit AD* value is representing. I know that lower is better fit but what is it.

0October 21, 2004 at 4:57 pm #109512You can try several options, some of them obviuos but it’s worth mentioning:

1. if your non normal data is composed by rational subgroups you can run the process capability for normal data (for example using Minitab)

2. you should see if the lack of normality is due a small sample which is not representative of your process

3. you can check if the lack of normality is caused by outliers

4. you can see if you have a multimodal distribution that can be splitted in 2 main groups each of them normally distributed

5. if you data is really non normal and follows a typical distribution you could try to use Crystal Ball to calculate the process capability

6. you can try to transform the data with the Box-Cox transformation to’force’ the normality

7. you can treat your data as discrete simply counting how many of them are outside the spec limits

8. you can use indicators like Cpk substituting the precentiles 99.865 and 0.135 to the standard deviation

0October 26, 2004 at 1:46 pm #109733What is this logistics curve. is is a method to transform data?

suneel0October 26, 2004 at 3:02 pm #109736Rob, sounds like your using minitab 14? If so you went to individual distribution identification under Quality tools, found the curve that yours most represents and then went to the Non Normal Capability analysis section and didn’t see your curve on the drop down? That happens alot with my students, what they don’t realize is they can scroll up on that drop down and there are more distributions. All of the distributions in the individual Id section are in the non normal capability section.

If you weren’t using minitab then download the free minitab 14 trial and crank it out. It’s easy and a great feature.0October 28, 2004 at 11:47 am #109873https://www.isixsigma.com/library/content/c020121a.asp

This article was very insightful for me.

0October 28, 2004 at 2:02 pm #109885

Jonathon AndellParticipant@Jonathon-Andell**Include @Jonathon-Andell in your post and this person will**

be notified via email.AD is a good tool, but you don’t want to stop there. Here’s a sequence that might work:

Create probability plots and see which model creates the best straight line. Be on the lookout for multi-modal distributions, which manifest themselves as parrallel segments.

Always plot the data in their time sequence, to see if they truly come froma single, stable process.

If you still are satisfied that a non-normal distribution applies, consider determining the probability of being out of spec, using the new distribution as a model.

Convert that probability back to an “equivalent” Z value.

I am not a big fan of blind transformations. Data follow non-normal distributions for legitimate reasons. Undertsanding the underlying possible processes that yields non-normal outcomes adds to our process knowldge, which is th ultimate purpose of all this effort.

Hope that helps!0November 1, 2004 at 2:38 pm #110076Build a confidence interval on the parameter, based on the distribution of your data. It can be constructed for any distribution, the normal is just the most documented test. Process capability will seem pretty straight forward once you build your confidence interval… make sure you understand the percentiles obtained from your distribution fit.

0 - AuthorPosts

The forum ‘General’ is closed to new topics and replies.