P-Value less than 0.05. but everything else in control

O

OmarEn

Hello guys,
I have a question about a Cpk output.

My two control charts (Mean/Range) are in control. No points beyond control limits, my histogram looks well, similar to a normal bell, symmetrical. Cpk and Ppk indices meet my customer's expectations too.

But P-value is less than 0.05. How can I explain to management that even when everything else seems to be OK, the capability study actually fails? and more important... Their question would be: What part of the process is necessary to fix if Cpk, Ppk, control charts, even distribution are OK?

Thanks all for your time in answering my question.
Best Regards.
 

optomist1

A Sea of Statistics
Super Moderator
Hi OmarEn,

Can you share/attach screen shots of control charts and of capability analysis output? What software did you use? Some detail re: the process, etc.?

Thanks, optomist1
 

bobdoering

Stop X-bar/R Madness!!
Trusted Information Resource
Also, what is the process and the process parameter your are looking at? Can you post the data you analyzed in an excel spreadsheet? Time ordered data is critical to capability analysis.
 
O

OmarEn

Also, what is the process and the process parameter your are looking at? Can you post the data you analyzed in an excel spreadsheet? Time ordered data is critical to capability analysis.
Guys,
Thank you for your time. Attached is the chart that Minitab gives me, as well as the excel spreadsheet with the data analyzed.

This is for a cutter tube machine and measurements were taken with a high caliper.

Thank you again for your help.
 

Attachments

  • Cpk.PNG
    Cpk.PNG
    23.9 KB · Views: 739
  • CPK Cortadora Haven T1 Brake LD.xlsx
    11.4 KB · Views: 338

stevegyro

Involved In Discussions
Please help with little more information sir:

1. Are you measuring length (‘Longitud’)?
2. Is ‘high Caliper’ max length, from several length measurements then recording the longest?

3. P value, when using MiniTab, usually asks for the ‘amount of difference’ considered significant (and uses this when comparing the mean of your sample, with the nominal).
Possibly a value (too small) was chosen, so it (MiniTab) says ‘yes the mean of your sample, is significantly different than the nominal.
I would not “hang my hat” or worry about this conclusion!

Low P values are often misunderstood. Their power increases if you are checking a one sided test, etc.

In reality, P values and the central limit theorem PRESUME you have “random & independent” sampling.
Using a cutoff tool does not create a normal (Gaussian) distribution, because the cutting tool is either changed, or ‘dressed’ at every (n) parts right?





Sent from my iPad using Tapatalk
 

Miner

Forum Moderator
Leader
Admin
The low p-value means it failed the test for normality. The probable reason for that is an unstable process. Try turning on some of the supplemental tests for out of control conditions, and I think you will find that the xbar chart is drifting.
 
O

OmarEn

Please help with little more information sir:

1. Are you measuring length (‘Longitud’)?
2. Is ‘high Caliper’ max length, from several length measurements then recording the longest?

3. P value, when using MiniTab, usually asks for the ‘amount of difference’ considered significant (and uses this when comparing the mean of your sample, with the nominal).
Possibly a value (too small) was chosen, so it (MiniTab) says ‘yes the mean of your sample, is significantly different than the nominal.
I would not “hang my hat” or worry about this conclusion!

Low P values are often misunderstood. Their power increases if you are checking a one sided test, etc.

In reality, P values and the central limit theorem PRESUME you have “random & independent” sampling.
Using a cutoff tool does not create a normal (Gaussian) distribution, because the cutting tool is either changed, or ‘dressed’ at every (n) parts right?





Sent from my iPad using Tapatalk
1. Yes, I am measuring longitude of a cut tube.
2. I just set the caliper zero at granite surface plate, then put the tube vertically and hit at top of it and capture the measurement.
3. The problem is that my customer won't accept this study if P-value is less than 0.05.
 

stevegyro

Involved In Discussions
OK sir.

The customer’s concern may need clarification, which is not always easy.

The histogram you shared with this thread is most helpful.
Notice how it is skewed a little to the right of the median (central measurement between highest and lowest).

This is no big deal. Are you using subgroups, IE does each mean and range point actually represent 25 pieces?
This is very important.

The Pp is better than the Cp, and as a supplier QE this would concern me. It is saying the variation ACROSS all subgroups (ie overall variation) is better (less) than the variation within subgroups. This is not usual.
Usually the overall Pp is worse (more variation) because it is over a longer qty of testing and time, so long term drift is added to the short term drift within groups.

Search a video about Ppk vs. Cpk BY MR. DENTON BRAMWELL.

He gives a great, patient discussion with real examples!

I believe he is right, and his key claim is that it is good to see Cpk = Ppk, and also Cp=Pp.

Your data is not bad at all!

Normality is very complicated. Please don’t fall for the P value trap. Machining or cutting metal is not random. Tool wear, temperature, and other factors make certain that it is not a purely random result.

If your client pushes too hard without giving good reason, maybe ask him if a GR&R would put him at ease.

If your process skews slightly, but your Process Performance and Capability are both over four, I would go out and celebrate this personally!

Cutting tools and abrasives/ abrasive wheels DO WEAR. The results felt in the cutting process do not follow a Central Tendancy.

Your client needs to understand this. I used to make the same silly oversimplifications (about P value).

Ask an expert in statistics to explain the internals of the Anderson Darling test, or Shapiro-Wilk or any from several others, and any of us will get dizzy from the logic.

Just look at the distribution!
Your data looks fine!

Look up ‘HiLo Mean and Range Charts, covered on “CorrectSPC Video Training” on the ASQ website.
Share that link with your customer!

Please let me know if you have subgroup (within) size > 1.
Thank you. Steve Giarratana







Sent from my iPad using Tapatalk
 

bobdoering

Stop X-bar/R Madness!!
Trusted Information Resource
To reject a capability study because the normal curve is not the best model is only applicable if:

1. The process variation is supposed to be normal - with random and independent process output (not always the case) and
2. The process has only one normal variation factor (nearly impossible - see total variance equation)

Very few processes actually meet that criteria, although significant gage or measurement error can make them appear to.

For your customer to assume that is the case for all processes and characteristics is disturbing and incredible misuse of statistics.
 

stevegyro

Involved In Discussions
To reject a capability study because the normal curve is not the best model is only applicable if:



1. The process variation is supposed to be normal - with random and independent process output (not always the case) and

2. The process has only one normal variation factor (nearly impossible - see total variance equation)



Very few processes actually meet that criteria, although significant gage or measurement error can make them appear to.



For your customer to assume that is the case for all processes and characteristics is disturbing and incredible misuse of statistics.



What Mr. Bob Doering says!

Could not have said it better. My favorite is his point that their are always several factors, but we usually do not ‘see’ them relative to other more significant factors, such as people, gage error, sampling error.

Great points!




Sent from my iPad using Tapatalk
 
Top Bottom