Dear friends,
How do you identifey Special causes in a process?
Have looked at TS/QS refmanual "SPS" 2nd edition page 75 have a criteria table, is this a way to get your process identified? Is there any moore information about this somewhere?

Is there any software that have this criteria table today?

Thx

Dear friends,
How do you identifey Special causes in a process?
Have looked at TS/QS refmanual "SPS" 2nd edition page 75 have a criteria table, is this a way to get your process identified? Is there any moore information about this somewhere?

Is there any software that have this criteria table today?

Thx
Special cause is defined as any observation that is outside the expected 3s limits of the process. Assuming a normal distribution, the probability of such an observation by chance alone is approximately 3 in 1000. Therefore, it is considered a special cause and not common cause variability.

The most common method for identifying special causes is a control chart.

Does this help?

Special cause is defined as any observation that is outside the expected 3s limits of the process. [.

That's +/- 3s limits, and any point beyond the limits should be a candidate for assignable cause, not a sure thing.

Assuming a normal distribution, the probability of such an observation by chance alone is approximately 3 in 1000.

Which means that 3 times in 1000 opportunities, a process free from assignable cause variation should be expected to have a point outside the limits.

Therefore, it is considered a special cause and not common cause variability.

It should be considered suspect and nothing more until further investigation is done.

There are also other probability-based tests for normality in control charts; this thread (http://elsmar.com/Forums/showthread.php?t=13866&highlight=tests+for+normality) might be helpful, especially the attachment to the first post.

Special cause is defined as any observation that is outside the expected 3s limits of the process. Assuming a normal distribution, the probability of such an observation by chance alone is approximately 3 in 1000. Therefore, it is considered a special cause and not common cause variability.

The most common method for identifying special causes is a control chart.

Does this help?
Have we helped answer the question or just defined the problem. I accept the control chart is probably the first indication that a "special cause" of variation exists but the means of "identifying" special cause is good old detective work - fishbones, five whys etc.

Steve, a point outside the 3 sigma limits is only one (potential - as Jim points out) indicator of special causes. There are many other indicators - as any good SPC book will describe - such as trends, mean shifts, cycles, etc.

Generally speaking, a good term that is often used in place of "special" is "assignable". Any point or pattern on a control chart that appears to be something other than random, is a candidate for a special, assignable cause, meaning that some unique and unusual thing has happened. An investigation is needed to determine what happened.

SPC books provide some rules, however (e.g. 7 successive upward/downward points), to help keep us from "chasing shadows", that are statistically sound.

niotusen, since it seems you have access to reference manuals from AIAG, please refer to their SPC Manual, 2nd ed., starting on page 69.

Dear friends,
How do you identifey Special causes in a process?
Have looked at TS/QS refmanual "SPS" 2nd edition page 75 have a criteria table, is this a way to get your process identified? Is there any moore information about this somewhere?

Is there any software that have this criteria table today?

Thx

There are many commercial SPC packages that will "flag" potential assignable cause variation--in fact, most will.

Thanx,
I did reffer to the AiAG SPC Manual Defining out of Control signals,

If the process is out of control. I had to check thru the crieterias, and define my process. Can I use own made criterias here?

Thinking.....
If my process is out of control should I take the a ppk calculation and not the cpk calculation in process??

Im a little confusion about this...

Is a PPAP study always calculated in PPK and when running the process after the PPPAP etc. I will calculate the Cpk?

Why I m asking?

My customer always ask a Cpk in our PPAP process, (Cpk 1.67 in PPAP and Cpk1.33 later on when we runs the process)

I´m confused here is the customer really asking for Ppk in this moment(PPAP)?

-----------------
Thx

Thanx,
I did reffer to the AiAG SPC Manual Defining out of Control signals,

If the process is out of control. I had to check thru the crieterias, and define my process. Can I use own made criterias here?

Thinking.....
If my process is out of control should I take the a ppk calculation and not the cpk calculation in process??

Im a little confusion about this...

Is a PPAP study always calculated in PPK and when running the process after the PPPAP etc. I will calculate the Cpk?

Why I m asking?

My customer always ask a Cpk in our PPAP process, (Cpk 1.67 in PPAP and Cpk1.33 later on when we runs the process)

I´m confused here is the customer really asking for Ppk in this moment(PPAP)?

-----------------
Thx

The AIAG PPAP manual (currently 3rd Edition, but scheduled for update on March 1) is where the PPAP requirements for "initial process studies" are found. The manual says (pages 6-7),

When historical data is available or enough initial data exist to plot a control chart (at least 100 individual samples), Cpk can be calculated when the process is stable. For chronically unstable processes withoutput meeting specifications and a predictable pattern, Ppk should be used.

I doubt that your customer wants and intial Cpk of 1.67 and then for you to coerce the process into deteriorating to 1.33. It's more likely that an intial 1.67 Ppk is expected, and the ongoing Cpk must be 1.33 or higher.

Slightly :topic: , but the quotation above contains what seems to be an irresolvable contradiction; an "unstable" process will not, by definition, exhibit a "predictable pattern."

I quess my customer want a ppk 1.67 and cpk 1,33 later on , but i our specs from the customer is in ppap Cpk 1.67 and Cpk 1.33 later ongoing?
What should I doo? Tell them they are wrong? Or have I missunderstood something here?
Thx

I quess my customer want a ppk 1.67 and cpk 1,33 later on , but i our specs from the customer is in ppap Cpk 1.67 and Cpk 1.33 later ongoing?
What should I doo? Tell them they are wrong? Or have I missunderstood something here?
Thx

I think it might be a good idea to discuss it with the customer, pointing out the fact that they're asking you to deteriorate the process.

Maybe, what your customer thinks (I am assuming this) that in a "more controlled situation" during PPAP you should get a bigger cpk than during the normal production of your parts. Just in case he isn't confused about the difference between Ppk and Cpk.

Yes they did not know the differences, but is there a possibility to have both indexes in a PPAP or is it always question of Ppk in PPAP?

Meening we have to produce 125 pcs to a PPAP and in my eyes there are no way to calculate a Cpk if we take the parts of the process all together, (Not in subgroups ex 5pcs as in calc of Cpk?)

(Did not have the PPAP manual in my hands at the moment therefore the quote)
Thx

It depends on the calculation of your standard deviation

Yes they did not know the differences, but is there a possibility to have both indexes in a PPAP or is it always question of Ppk in PPAP?

If your customer requires PPAP in accordance with the AIAG requirements, then the requirements in the current PPAP manual apply unless your customer has specifically superseded them.

Meening we have to produce 125 pcs to a PPAP and in my eyes there are no way to calculate a Cpk if we take the parts of the process all together, (Not in subgroups ex 5pcs as in calc of Cpk?)

The default PPAP requirement (at this moment) is for x-bar/R charting during the "significant production run," with a minimum of 25 subgroups and 100 individual data points. Capability index requirements were quoted earlier. If you haven't done the study according to the requirements, and your customer hasn't approved a deviation from the requirements, then you haven't met the requirements.

The problem is,
if we make 100 (think it was 125) individual points should this be calculated with Ppk and the 25 subgroups will be calculate as Cpk? We need to show both?

or is this the same data points, meening 100 individual point with a minimum of 25 subgroups taken of the 100 points of our ongoing process?

Think I am a little way out here?
We have only make earlier 125 data points and calculate Cpk of 125 data points. This must be wrong....

The problem is,
if we make 100 (think it was 125) individual points should this be calculated with Ppk and the 25 subgroups will be calculate as Cpk? We need to show both?

or is this the same data points, meening 100 individual point with a minimum of 25 subgroups taken of the 100 points of our ongoing process?

Think I am a little way out here?
We have only make earlier 125 data points and calculate Cpk of 125 data points. This must be wrong....

The difference between Cpk and Ppk is that the former uses the sample standard deviation (based on within-subgroup variation) and the latter uses the population standard deviation. If you have logical subgroups and haven't measured every member of the population, you should use Cpk. You should not arbitrarily form subgroups from a general population; the samples must be selected at prescribed intervals while production is running. On the other hand, if you just have a lot of 125 parts and you're going to measure each one, you should use Ppk, because the standard deviation calculation doesn't assume sampling and subgroups.

Whatever you do, make sure that the customer understands and approves, and it makes good statistical sense.

Thanks I think we will be online again, but the strange thing is that the customers doesn´t know what they want from us as a supplier and they never heard about differents of Ppk and Cpk calculation....
And the best way is if the process are in order we use Cpk and when out of order we use Ppk,
but often we dont have enought time to adjust and run the process before the parts has been made for PPAP the process isn´t become stable then we must calculate the ppk value.
What the customer wants must deepend on how the process results has been a in order or out of order process?
Thanks and regards from Sweden..

I had a quick side comment/question about nomenclature (specifically for Jim but far any other interesed parties as well).

To a statistician, "sample standard deviation" (STDEV in Excel) and "population standard deviation" (STDEVP in Excel) mean two specific things.

STDEV = [ Σ(x - x-bar)^2 / (n-1) ]^0.5
STDEVP = [ Σ(x - x-bar)^2 / (n) ]^0.5

Suppose you collect 4 consecutive parts every hr for 30 hr.

As I understand it, Cpk is based on the st dev of the parts within subgroups. Often this is estimated from the ranges (Max - Min for each set of 4) of each of the 30 subgroup. It could also be estimated from the sample st dev (STDEV of the 4 parts). You look up the magic numbers from the control chart tables and :magic:you get an estimate of the standard deviation within the sets of 4.

Ppk, OTOH, is based on the STDEV of the entire set of 120 parts lumped together. To a statistician, this is still a "sample standard deviation". (If the 120 parts represent the entire run, then you should probably use STDEVP, but for large sets like this it doesn't matter much whether you divide by n or n-1.)

If the variation within the sets accounts for all the variation (ie there is no other long-term variation), then the two estimates for the standard deviation will be the same, and Cpk = Ppk.

If the variation within the sets does not account for all the variation (ie there is some other long-term variation), then the standard deviation estimated from the sets of 4 will be less than the standard deviation estimated for the whole set of 120, and Cpk > Ppk.

Although there are two ways to estimate the standard deviation (and even though one method uses small samples and the other uses the combination of all the samples), it is not really a question of "sample standard deviation" vs. "population standard deviation". It is more a question of standard deviation of parts within subgroups vs standard deviation of all the parts sampled.

Is that the way others see it? Is that a reasonable interpretation of Cpk & Ppk?

Tim F

I had a quick side comment/question about nomenclature (specifically for Jim but far any other interesed parties as well).

To a statistician, "sample standard deviation" (STDEV in Excel) and "population standard deviation" (STDEVP in Excel) mean two specific things.

STDEV = [ Σ(x - x-bar)^2 / (n-1) ]^0.5
STDEVP = [ Σ(x - x-bar)^2 / (n) ]^0.5

Suppose you collect 4 consecutive parts every hr for 30 hr.

As I understand it, Cpk is based on the st dev of the parts within subgroups. Often this is estimated from the ranges (Max - Min for each set of 4) of each of the 30 subgroup. It could also be estimated from the sample st dev (STDEV of the 4 parts). You look up the magic numbers from the control chart tables and :magic:you get an estimate of the standard deviation within the sets of 4.

Ppk, OTOH, is based on the STDEV of the entire set of 120 parts lumped together. To a statistician, this is still a "sample standard deviation". (If the 120 parts represent the entire run, then you should probably use STDEVP, but for large sets like this it doesn't matter much whether you divide by n or n-1.)

If the variation within the sets accounts for all the variation (ie there is no other long-term variation), then the two estimates for the standard deviation will be the same, and Cpk = Ppk.

If the variation within the sets does not account for all the variation (ie there is some other long-term variation), then the standard deviation estimated from the sets of 4 will be less than the standard deviation estimated for the whole set of 120, and Cpk > Ppk.

Although there are two ways to estimate the standard deviation (and even though one method uses small samples and the other uses the combination of all the samples), it is not really a question of "sample standard deviation" vs. "population standard deviation". It is more a question of standard deviation of parts within subgroups vs standard deviation of all the parts sampled.

Is that the way others see it? Is that a reasonable interpretation of Cpk & Ppk?

Tim F

I don't see the distinction you're making. Perhaps sample and population are misleading, but I didn't give them those names.

I don't see the distinction you're making. Perhaps sample and population are misleading, but I didn't give them those names.
Let me give you a different set of nomenclature to try on for size.

Cpk uses the within run or repeatability estimate of variance.

Ppk uses the total variance that includes within run and run to run estimates.

Estimation of standard deviation (or variance) by lumping the 100 points together is wrong! You need to use variance components to get the within run and run-to-run components.

Let me give you a different set of nomenclature to try on for size.

Cpk uses the within run or repeatability estimate of variance.

Ppk uses the total variance that includes within run and run to run estimates.

Estimation of standard deviation (or variance) by lumping the 100 points together is wrong! You need to use variance components to get the within run and run-to-run components.

Isn't that what I said?

Isn't that what I said?
Jim, you did indeed say the same thing, but I got the impression people would just take the sd of all the data and not use variance components.

Jim, you did indeed say the same thing, but I got the impression people would just take the sd of all the data and not use variance components.

I understand. I'm glad you and Tim are here to help clear these things up.

Just to be persnickety, I'll add a few coals to the fire.

If a newbie reading this thread is a student, studying from a textbook, then what Tim says is correct and not misleading. A SAMPLE standard deviation is what is used most often, since the total POPULATION is more often unknown. So the sample standard deviation calculation factors out bias (using n-1). It made perfect sense to me.

So:
- a small known population will use population standard deviation (but then again, if that is the case, what's the point?)
- a large discrete population from which you take a sample uses sample standard devation (and this is what short term capability studies are),
- a large "running" population from which you take samples periodically uses an estimate of sample standard deviation, like R bar/d2 (and this is what long term capability studies use).

Steve, be careful mixing terms like variance and standard deviation without explaining the difference. Newbies might get confused, since variance is the square of standard deviation, not another word for it. Also, using the word repeatability, which is more often used for measuring device integrity than in SPC, since measures of dispersion are related to consistency, not repeatability.

Of course, all of this is just for us anal-retentive types who get all excited seeing a mean shift on an X-R chart! :rolleyes:

Thank for all..
I have been in discussion with other QA people and it seemes that many are doing PPAP a little different ways..

Many take out samples of 100 and calculate the ppk. (some even get the datapoints to subgroups and calculate a Cpk of that.)

It seemed they never take out 25 soubgroups in Xbar R diagram and then calculate Cpk of that in the right way that I see it?

How are your reflexions/thinkings of this?

Just to be persnickety, I'll add a few coals to the fire.

If a newbie reading this thread is a student, studying from a textbook, then what Tim says is correct and not misleading. A SAMPLE standard deviation is what is used most often, since the total POPULATION is more often unknown. So the sample standard deviation calculation factors out bias (using n-1). It made perfect sense to me.

So:
- a small known population will use population standard deviation (but then again, if that is the case, what's the point?)
- a large discrete population from which you take a sample uses sample standard devation (and this is what short term capability studies are),
- a large "running" population from which you take samples periodically uses an estimate of sample standard deviation, like R bar/d2 (and this is what long term capability studies use).

Steve, be careful mixing terms like variance and standard deviation without explaining the difference. Newbies might get confused, since variance is the square of standard deviation, not another word for it. Also, using the word repeatability, which is more often used for measuring device integrity than in SPC, since measures of dispersion are related to consistency, not repeatability.

Of course, all of this is just for us anal-retentive types who get all excited seeing a mean shift on an X-R chart! :rolleyes:

Thanks for the added clarification, Rob:agree1:

Thank for all..
I have been in discussion with other QA people and it seemes that many are doing PPAP a little different ways..

Many take out samples of 100 and calculate the ppk. (some even get the datapoints to subgroups and calculate a Cpk of that.)

It seemed they never take out 25 soubgroups in Xbar R diagram and then calculate Cpk of that in the right way that I see it?

How are your reflexions/thinkings of this?

I see a lot of goofy stuff in PPAPs, niotusen. You can never go wrong by adhering to the requirements, though, regardless of what others are doing. You might just make a favorable impression on your customer if you do things the right way while everyone else is doing something else.

Just to be persnickety, I'll add a few coals to the fire.
Steve, be careful mixing terms like variance and standard deviation without explaining the difference. Newbies might get confused, since variance is the square of standard deviation, not another word for it. Also, using the word repeatability, which is more often used for measuring device integrity than in SPC, since measures of dispersion are related to consistency, not repeatability.

Of course, all of this is just for us anal-retentive types who get all excited seeing a mean shift on an X-R chart! :rolleyes:
Rob, I agree I should be clearer on differences between variance and standard deviation, but I assumed (I know!...) people who are 6S experts would know the difference.

I disagree with your assessment that repeatability is only used for measurement devices. I use repeatability as an equivalent to within run and reproducibility as run to run. If you rather use consistency, then which source are you referring to?

You are not incorrect Steven. Juran's Quality Handbook provides a definition very much like yours, i.e.,In manufacturing, reproducibility measures the variability between items manufactured on different days or on different machines. Repeatability measures sources of variability that are more local or immediate, assignable to item measurements or to the variability occurring between adjacent items manufactured in sequenceBut interestingly, that description is given on the section dealing with Design of Experiment and is never mentioned in the "SPC" section. There, it only uses the terms, "within group variation" and "between group variation" - which is how I prefer to spell it out.

Where I work we build machines, and therefore do a lot of both GR&R and short term capability studies (Ppks), so I don't want to confuse the people with the same nomenclature for, as an example: 1) one part checked five times, and 2) 5 parts checked sequentially.

Where you work, as long as everyone understands the terms the same way, I see no problems. :agree1:

How does the sample size affect things in the case of trying to determine distribution of a compositon.

For example if my process produces a 1000 lbs of material that is a mixture of two solids. The two solids are the same size and shape and weight (lets call them beads) but different colors. The weight of a single bead is very small compared to the overall weight of the mixtrue let say 0.01 lbs per bead. The mixture should be 50:50. You repeate this 10 times for a total of 10,000 lbs. Each batch goes into a single drum. How many samples do I need from each drum to calculate capability. AND more important what size samples should I use, 1 lb, 0.1 lb, ect.

If I completly analyse each drum I would expect a very small standard deviation, (my scales used to fill the drums are very accurate and mixing in the drum is not a factor because I measured the whole thing). But as I test smaller and smaller samples the distribution of the beads within the drum becomes a larger factor. So I seems I can get any Cpk I want by changing the sample size.

So how to deal with this?

Should the sample I take be based on how the material is used? For example if I injection mold a part from this blend should my sample be ~the same weight as a single part? It seems like this would be correct. Can anyone give me some statistical details to back this up?

What If my measurement tool can only handle a very very small sample. For example my test is DSC which might use a few miligrams of material. And my part weight is 50 grams. How do I calculate how many 1/10th of gram samples to measure in my original mixture?

Thanks for any insight.

whather ur talking about assignable & unassignable causes of variation
than use control chart
its a best method