Informational Control Chart Interpretation - General "Rules"

Jim Wynne

Leader
Admin
Lately I find myself spending a lot of time explaining the basic "rules" of control chart evaluation to suppliers and coworkers. I put together the attachment to use as a basic explanation and reference, mostly so I could stop having to repeat the same explanation over and over again. I wanted something reasonably concise--a single page--and wasn't attempting a treatise. I'd like to have my fellow Covers look it over and offer suggestions for improvement, and also offer the reference to anyone who might find it useful.

Note: the attachment has been updated to correct an error pointed out by Tim Folkerts--see his post and my response to it in this thread. Thanks, Tim :bigwave:
 

Attachments

  • Control Charts - General Rules for Interpretation.doc
    15.5 KB · Views: 2,593
Last edited by a moderator:

Statistical Steven

Statistician
Leader
Super Moderator
I find that the 2 out 3 points outside 2S and 4 out of 5 outside 1S to be overkill. I stick with outside 3S and trends either up or down. Do you find people understand how to set up control charts? Most importantly, how much data is needed to set the initial limits, and when to recalculate limits?
 

Jim Wynne

Leader
Admin
Statistical Steven said:
I find that the 2 out 3 points outside 2S and 4 out of 5 outside 1S to be overkill. I stick with outside 3S and trends either up or down.

I appreciate the input. Actually, the phenomena you refer to are just as unlikely (and perhaps more so) to happen as points beyond the limits. I want people to be aware that they need to look at the process when anything significantly improbable happens.

Statistical Steven said:
Do you find people understand how to set up control charts? Most importantly, how much data is needed to set the initial limits, and when to recalculate limits?

I'm sorry to say that anything beyond a very superficial understanding of basic normal-curve statistics is rare, in my experience.
 
B

Bill Ryan - 2007

:agree1:
Jim I like it (and may (probably will) "borrow" from it as we approach our company wide training for SPC). It doesn't have "too much" math principle in it (which tends to "lose" the trainee :confused: ). If someone asks for a more in-depth explanation and really wants to understand the underlying principles there are many resources to point them to.
 

Jim Wynne

Leader
Admin
Bill Ryan said:
:agree1:
Jim I like it (and may (probably will) "borrow" from it as we approach our company wide training for SPC). It doesn't have "too much" math principle in it (which tends to "lose" the trainee :confused: ). If someone asks for a more in-depth explanation and really wants to understand the underlying principles there are many resources to point them to.

Thanks, Bill. I'm glad to see it being put to good use. I avoided the 'rithmetic on purpose, for the reason you mention. It's a very deep subject, and I just wanted to emphasize a few basics.
 

Tim Folkerts

Trusted Information Resource
Overall I like the approach. I do have a few specific comments, though.

You said.
"Cpk should not be calculated unless the process is in a state of statistical control, (because the estimate sigma of used in the calculation will be unreliable)"
To me, the primary reason to require control has to do with predictability. If the process is not in control, then you don't have a good understanding of how it was behaving in the past. If you don't know how it was behaving, then you sure as heck don't know how it will behave in the future!

"Several tests for normality have been developed..."
More specificially, I think you mean "tests for stability" or "test for control". You are not actually testing the data to see if it follows a normal distribution, which is the literal interpretation of this phrase.

"Because we know that in a normal distribution 99.73% of the population will fall within ± 3s of the mean, we also know that there is only a .27% chance that a point will fall outside those limits."
Many processes do not follow a normal distribution, so these specific numbers are not always a good estimate of actual probabilities. I would suggest being more generic, like "Because we know that the vast majority of the data must fall within ± 3s..."

"In other words, there is a 99.73% chance that the points outside the limits are the result of non-random causes."

That doesn't quite follow. If the process is truly in control and following a normal distribution, then there is a 0% that the points outside the limits are the result of non-random causes. You could drop this sentence without really losing the point you are making.

Just my $0.02...

Tim
 

Jim Wynne

Leader
Admin
Tim Folkerts said:
You said.
"Cpk should not be calculated unless the process is in a state of statistical control, (because the estimate sigma of used in the calculation will be unreliable)"
To me, the primary reason to require control has to do with predictability. If the process is not in control, then you don't have a good understanding of how it was behaving in the past. If you don't know how it was behaving, then you sure as heck don't know how it will behave in the future!

A good observation, but my point was that at the time that a rule violation is discovered, it's too early to conclude whether or not the process is stable, but the calculation shouldn't be done because the integrity of the estimate of sigma is in question.

Tim Folkerts said:
"Several tests for normality have been developed..."
More specificially, I think you mean "tests for stability" or "test for control". You are not actually testing the data to see if it follows a normal distribution, which is the literal interpretation of this phrase.

Yes--I should have said "stability" instead of "normality.":eek: It's an important distinction and I'll correct the original

Tim Folkerts said:
"Because we know that in a normal distribution 99.73% of the population will fall within ± 3s of the mean, we also know that there is only a .27% chance that a point will fall outside those limits."
Many processes do not follow a normal distribution, so these specific numbers are not always a good estimate of actual probabilities. I would suggest being more generic, like "Because we know that the vast majority of the data must fall within ± 3s..."

This is intended to be general, and my point about probability is true in a general sense, although your point regarding non-normal distributions is well taken. Believe me though, we have to teach them to walk before we can expect them to run:) .



Tim Folkerts said:
In other words, there is a 99.73% chance that the points outside the limits are the result of non-random causes."
That doesn't quite follow. If the process is truly in control and following a normal distribution, then there is a 0% that the points outside the limits are the result of non-random causes. You could drop this sentence without really losing the point you are making.

The hypothesis is that 99.73 percent of a normally distributed process will fall between +/- 3 standard deviations of the mean. That means that .27% of a normally distributed process will fall outside of the three-sigma limits. Thus it's possible for the process to be stable and still produce points outside the limits.

Thanks, Tim. I really appreciate the observations:agree1:
 

Marc

Fully vaccinated are you?
Leader
Also see Interpreting Control Charts - You might want to include a link to people you're trying to 'educate'. Everyone has computer access any more - You never know, some people may want more details.

The powerpoint file is 'incomplete' - Long story...
 

Caster

An Early Cover
Trusted Information Resource
Non math approach to SPC

JSW05 said:
Lately I find myself spending a lot of time explaining the basic "rules" of control chart evaluation to suppliers and coworkers. I put together the attachment to use as a basic explanation and reference, mostly so I could stop having to repeat the same explanation over and over again. I wanted something reasonably concise--a single page--and wasn't attempting a treatise. I'd like to have my fellow Covers look it over and offer suggestions for improvement, and also offer the reference to anyone who might find it useful.Thanks, Tim :bigwave:

I like it a lot.

One thing I have now done here is to use only real examples from our own shop. We now have lots of examples. I started with the examples fropm textbooks, buy now use only our charts.

We only have 3 rules in play (points out of control, sudden shifts, and mixture), so that's all I explain.

I also teach people to look at the range chart first....only because most people ignore it.

I also don't talk in terms of probability and the normal curve...instead I give everyone 2 dice and have people roll them and shout out the numbers. I record the results in a Pareto and after about 80 throws, or less than 1 minute I have a nice bell shaped curve.

This 2 dice machine makes numbers...from 2 to 12...it mostly makes 7's....if it makes a 1 it is broken...and my favorite (per Deming) the operator is not to blame or in control....I usually tell people that I will bankroll anyone at the Casino if they can control the outcome of this machine!

Then I say the control limits are set "near" the ends of the curve and act as your signal to look at your machine, because you have an unexpected result....everyone gets this and nods their head...then I offer my time to go through the math if anyone wants it.....a few people do take me up.

People also get the idea of a run easily....if someone start rolling all 7's....you need to shoot them....they're cheating!

So there's my no math approach to it all. Don't get me wrong, I love the math! But no one needs any math to get SPC.

Also as per Shewart, Wheeler, and Prevette, normality is not required for SPC! (I know, I know, it is required for prediction and capability.)
 
Top Bottom