Non-Normal Data & Minitab

[Mexicanquality]

Gentleman,

I have been lately working for new customer and they have been pushing me to use Minitab to calculate Ppk, Cpk, etc, but my data aren´t Normal data (anderson test) and because of this I have to find the best distribution (weibull, largest, etc) or to tranform the data with Jhonson or Box cox.
My main questions are:
-Why Minitab doesn´t calculate the Cpk and Cp on Non normal data (is because the formules are based in normal data??)
-In some tranformations Minitab doesn´t calculate Pp, is there any reason?
-What is next if no one distribution or transformation fits to data? we assume nornamilty?

I know are a lot of questions but I think all of them are related.

Joe

 

[AdamP]

Re: Non normal data & Minitab

Cpk and Cp are based on having normally distributed data.

If you can fit the data, conducting the appropriate capability study is pretty easy. If you have a spec limit (or 2) you can determine the percent outside that limit. You can then use the percent non-conforming to arrive at a 'per million' figure.

Cheers.

 

[bobdoering]

Re: Non normal data & Minitab

Gentleman,

I have been lately working for new customer and they have been pushing me to use Minitab to calculate Ppk, Cpk, etc, but my data aren't Normal data

If your customers are automotive, you may want to refer to PPAP 4th edition section 2.2.11.5 Processes With One-Sided Specifications or Non-Normal Distributions.

 

[Mexicanquality]

Re: Non normal data & Minitab

But what about if data doesn´t fit any distribution and aren´t possible to tranform? in this case should I use normal distrubution??? the manual doesn´t specify this.

Thanks

 

[bobdoering]

Re: Non normal data & Minitab

But what about if data doesn´t fit any distribution and aren´t possible to transform? in this case should I use normal distribution??? the manual doesn´t specify this.

You are correct, it does not specify much helpful information beyond normal. It really depends on the data and the source(s) of variation. Can you share the data? I really can not say what to do without seeing the data. If the normal distribution is not a meaningful model, you have no reason to assume using it is correct. Think in terms of the models weathermen use to predict the weather. If the model does not fit, the prediction will likely be useless.

 

[Statistical Steven]

Re: Non normal data & Minitab

But what about if data doesn´t fit any distribution and aren´t possible to tranform? in this case should I use normal distrubution??? the manual doesn´t specify this.

Thanks

What version are you using? Minitab does do Capability analysis for non-normal data.

 

[falconer65]

What type of process is it? What features? If I am stamping a hole in sheet metal it will be a non-normal distribution, but predictable.

Has a GR&R been done? If the measurement system is incapable of consistency, your data will be as well.

What is your sample size?

Are you using sub-groups?

These are the first that come to mind.

Erik

 

[Mexicanquality]

Re: Non normal data & Minitab

Minitab 15, I have used Weibull and largest but there are some data that doesn´t fit to any distribution and either weren´t able to transform with Johnson and box cox, in this case what should I do?

 

[Mexicanquality]

One of the features is a profile, the othe one a true position, the measuring system is the CMM (of course calibrated), should I run a GR&R gor CMM?

Whay did you mention that a stamp hole will be a non-normal data?

Sample size is 25 subgroups with 5 part per subgroup (125pzs)

Thanks for your time

 

[bobdoering]

It would be far easier to help if we could see that raw data. To many possible variables to come close to giving you a definitive answer. There needs to be time-sequence data plotted, along with the distributions.

The data by itself is not enough. There needs to be an understanding of the potential variation in the process. Multimodal variation - especially of different distributions - can confound simple analysis (like a capability calculation).

Remember, data supports or rejects the theory - it does not generate it.