Determining whether a process yeild is normal or non-normal distribution?
Hello!
Is there any recomendations of evaluating the process normal or non normal?
Any helpfull excel templates or small software to check the process distribution, stability etc. Capability etc is calculated in a software but no normality test?
Any ideas to test/evaluate the process?
We use Minitab, which is a great statistical product, for our capability studies... however, it's not free... I think we paid about $1000 USD for it a few years ago.
You can probably get 80% of the way there by plotting a histogram and visually comparing it to the standard normal curve. For example, it should have a single peak, centered at the mean, and the tails should be similar in size to each other. Many manufacturing processes are not normal, and often there are simple reasons. For example, if you have 2 or more process streams (e.g. 2 different dies or molds), you may get a bi-modal
distribution. If this occurs, you should do your SPC on the two dies or molds separately.
Another problem that may happen is that the distribution may be skewed, if for example it's not possible for a negative dimension to occur. We have this problem when we do SPC for flatness; it's not possible to have negative flatness.
If you're really determined, I think that you could use the chi squared distribution to test the normality of the distribution. The chi square
distribution is built into excel. I've never done this in Excel but it shouldn't
be that difficult.
It is well worth the effort to verify that a distribution is normal... most of SPC is based on the assumption of normality, and if the distribution is not normal,
you will be wasting your time and confusing yourself if you use statistics
when the assumption is not accurate.
Chi Square tests in excel are not too bad to pull off, though you will have to do a fair amount of coding yourself.
Another thing to consider is check if the "skewness" and "kurtosis" are both zero.
__________________
Steve Prevette
"A Passionate Statistician", ASQ CQE, Fluor Government Group
The opinion stated above does not necessarily reflect that of my employer.
SPC is distribution free and empirically based. Normality is not required.....Shewhart and Deming were clear on this point.
My 2 cents;
Chart your process, identify special causes, eliminate them and improve.
Rich DeRoeck
Rich, Steve, or Anyone who can Help,
I have been told that I am wasting my time on a U-Control Chart because we are trying to monitor a Process that is Not Normal, or at least our attempt to monitor the Order Quoting-Entry-Review-Releasing-Billing (QERRB) Process is not accurate enough. So I have been generating a Histogram with Descriptive Statistics to determine if the number of defects falls in a normal distribution or not. So I have been watching the P Value from the Anderson-Darling Normality Test, to see if and when it would go below .05. In the 42 weeks we have been monitoring these defects, the P Value has fallen from around .9 to .4. Am I way off in this effort? Please see attached file for the various worksheets.
You are correct, the U-chart is non-normal. It is in fact related to the Poisson. Do not take control charting as an exercise in probability. It is an empirical rule which has some good theory and testing behind it. I don't see what you are referring to with the .9 and .4 p values.
Looking at the U-chart, there is an out of control low point on week 33. It looks to the eye that something may have shifted in week 25. The effect of this shift may be hidden because you are updating the baseline with new data values. If you were happy with the initial baseline you should leave it alone.
What I would do is weeks 1 - 24 into a baseline and lock that baseline (this is based upon knowing that week 33 was low). Then I think you will see a lowered baseline for weeks 25 to 42, with the exception of a significantly high spike in week 34 (it is interesting it comes after a drop in week 33 - could some stuff have been counted in 34 that should have been counted in 33?)
The Pareto is a good idea, but be sure to do it only over a time interval that the control chart shows is stable.
Now, all this is based upon only looking at the chart. What happened with the process? Did something change around week 25?
__________________
Steve Prevette
"A Passionate Statistician", ASQ CQE, Fluor Government Group
The opinion stated above does not necessarily reflect that of my employer.
Do not take control charting as an exercise in probability. It is an empirical rule which has some good theory and testing behind it.
A control chart is nothing if not "an exercise in probability." What's the point in doing control charting if the charts have no predictive value? How can you have predictive value without a basis is statistical probability?
__________________
Some men are born mediocre, some men achieve mediocrity, and some men have mediocrity thrust upon them.-- Joseph Heller
A control chart is nothing if not "an exercise in probability." What's the point in doing control charting if the charts have no predictive value? How can you have predictive value without a basis is statistical probability?
This is a point that was made by Dr. Deming on several occasions. It has taken me a while (as a classically trained Operations Researcher) to get used to this idea, but I think it has merit. A lot of folks teach SPC using the normal distribution and making statements like 99.7% of the data will be between the control limits. There are several things which would cause the 99.7% to be incorrect - such as the data aren't normal (which they aren't in the u-chart) and the calculated baseline may be off (sample vs. reality) and the process data may not be stable. A concern for the 99.7% has sent several people off to mathematically transforming their data, or trying to calculate what the 99.7% limits would be for their distribution. This is not necessary.
Dr. Deming was also against the traditional uses of hypothesis testing.
But we do both agree (and so did Dr. Deming) that the control chart's purpose is prediction.
__________________
Steve Prevette
"A Passionate Statistician", ASQ CQE, Fluor Government Group
The opinion stated above does not necessarily reflect that of my employer.
Determining whether a process yeild is normal or non-normal distribution?
Linear Mode
