What statistical method to establish a specification?

[rmacd]

My company developed a characteristic and a destructive test for a purchased component. What statistical method would be appropriate to establish a specification for this characteristic using the test data?

Thanks in advance, Rod

 

[Statistical Steven]

That is a loaded question. I assume you have sufficient test data to know if the data is normally distributed. I assume the destructive test is sufficently discriminate to give a continuous measurement, and not just a pass/fail result.

Given that is the case, you can build a model to explain the variability (lot to lot and within lot and other sources of error). Using the different sources of error you can use mean+1.96S, where S is the sources of error you will experience when testing.

Again, you can consult your corporate statistician for help.

 

[Dave Dunn]

What is the characteristic and how does the characteristic affect the function of the assembly it goes into, or if it's a stand-alone part how does the characteristic affect the function of the part?

There must be a reason why you've decided to make this measurement. My assumption is that you're concerned that the part may fail under stress, whether in an assembly or by itself in use. Simply setting a specification based on what you're receiving for product currently may control the problem of failures, but without a logical reason that the specification should be at a particular level, you may end up rejecting parts that for all intents and purposes will function perfectly simply because they're not quite as good as what you started with.

For example: parts you receive are tested and sustain a torque force of 500 foot/pounds. After engineering experimentation, you determine that for your application, they need to sustain a minimum torque of 200 foot/pounds. Are you going to reject the more recent parts simply because they test at 400?

My suggestion is to get engineering involved and figure out what you need to make the part work without failure, then make sure your supplier can provide you with a process that is capable to make them.

 

[Bev D]

There are several statistical methods, but we need to understand the situation a bit more...what is this characteristic? an input factor or an output characteristic that can be measure non-destructively that you are attempting to correlate to a functional faliure (only found destructively)? is the data continous or categorical (pass/fail)

 

[Tom Slack]

My company developed a characteristic and a destructive test for a purchased component. What statistical method would be appropriate to establish a specification for this characteristic using the test data?

Thanks in advance, Rod

Specifications are defined in a statistical sense as "how bad can this part be and still meet be acceptable". The statistical tool of choice would be Design of Experiments.

I feel your question deals with the supplier's capability to produce acceptable and consistant parts. The tool of choice would be Process Capability Studies. There are tons of information on this web site and others on this subject.

Good Luck,

Tom

 

[Jim Wynne]

Specifications are defined in a statistical sense as "how bad can this part be and still meet be acceptable". The statistical tool of choice would be Design of Experiments.

A part isn't "bad" until it's bad . Variation is inevitable.

I feel your question deals with the supplier's capability to produce acceptable and consistant parts. The tool of choice would be Process Capability Studies. There are tons of information on this web site and others on this subject.

But how can you do process capability studies without specifications? Of course it's possible to run a process for the purpose of identifying the extent of inherent variation (which has to be taken into consideration, of course, when developing specifications), but the term "process capability" is generally construed to mean comparison of variation to spec limits.

 

[Statistical Steven]

Specifications are defined in a statistical sense as "how bad can this part be and still meet be acceptable". The statistical tool of choice would be Design of Experiments.

Actually DOE will help you understand sources of variability. If you just wanted to set a specifications that accepts 95% of the product, you could just do a 95% confidence interval about the mean. If you want to get fancy, you can use prediction intervals to determine what the mean of n items would be 95% of the time.

 

[Barbara B]

Maybe I'm too blind to see Aren't the specifications the limits which are given by customer or process requirements and not choosen in consideration to process output? (In contrast to control and warning limits which are calculated.)

You can of course establish specification limits based on test data, but what would you achieve? If the requirements were over-fulfilled, you get a too small range for the process. If the requirements couldn't be met by the actual process output, specs calculated out of the test data lead to complaints.

IMHO control limits and specs are statistically independent.

Barbara

 

[ralphsulser]

I remember years ago we were taught to use ANOVA and sum of squares, plus maybe chi square testing, to establish spec limits. I think it was in a data analysis and significance testing course. Anyone else remember this?

 

[Jim Wynne]

Maybe I'm too blind to see Aren't the specifications the limits which are given by customer or process requirements and not choosen in consideration to process output? (In contrast to control and warning limits which are calculated.)

What if the customer's specification limits can't be achieved? You assume too much if you think that customers' spec limits always take the laws of physics into account . On the other hand, the customer has to have some method of setting reasonable specification limits for new products, which must take into account the state of the art at the time the limits are set.

You can of course establish specification limits based on test data, but what would you achieve? If the requirements were over-fulfilled, you get a too small range for the process. If the requirements couldn't be met by the actual process output, specs calculated out of the test data lead to complaints.

Either test data or knowledge of the capability of a given process (i.e., the reasonably expected range of variation) must be used in order to set spec limits. How else can you know whether or not the specifications are reasonable?

IMHO control limits and specs are statistically independent.

They are statistically independent, but inextricably related. You can't have the latter without knowledge of the former.