Statistical Sampling

[racer12306]

I've searched around a bit, looked at MIL STD105E, etc, and am trying to come up with a reasonable sample size. But some of this stuff doesn't really apply. I'm not manufacturing so an acceptable quality level doesn't apply, and I still don't quite understand different inspection levels other than GII is the common one.

My situation:

I have a spreadsheet with 900 jobs on it. There is a rule that is supposed to be followed for each of these jobs, where applicable. Obviously adding a 900 line audit to my work load is not feasible, so I figure a smaller quantity monthly audit would be suitable for figuring out what I want to know. 10% is 90, still quite a bit, but if 10% is most reasonable I will find a way to make it work.

Is there a formula for figuring this out?

 

[Turk]

There's several ways you can do it and I only have a few minutes. One way:

You can take 35 sample measurements and create a Control Chart.

The Average is your mid-point (X-bar), then +/- 3 Standard Deviations can be your UCL (Upper Control Limit) and LCL (Lower Control Limit). Do your samples at determined intervals, depending on how critical, and plot them. If they fall out of the control limits, stop and solve the problem.

OR

You can do 95/95 Tolerance Intervals based on Juran's Quality Control Handbook. Use either the one-sided or two-sided chart to determine your k value.

 

[bepperb]

It's impossible to say, is it a pass/fail sort of thing or a value that you can compute standard deviations with?

 

[racer12306]

Thanks Turk



It's pretty much pass/fail. If the result is fail, I lecture someone about not doing stuff properly.

 

[wkcars]

Check out www.sqconline.com mil-std-1916 is the replacement for mil-std-105, if it's important or critical be sure to verify the statiatics yourself.

 

[KCJeep]

I use a sample size calculator we have on Excel, I don't really know where it comes from though.

A 95% confidence level is a common standard for statistical sampling. Using your 900 figure, the calculator result is 109 samples gives you a 95% confidence level + or - 3%.

If you wanted to drop to 80% you'd only have to do 30 though.

 

[Turk]

Originally Posted By: racer12306
Thanks Turk



It's pretty much pass/fail. If the result is fail, I lecture someone about not doing stuff properly.


Oh... Pass/Fail... To have a statistically significant Attribute process, you need 59 samples without fail. Variable needs at least 30.

Are you/did you perform a Validation, aka Performance Qualification (PQ) with EVT (Extreme Value Testing) High & Low?

 

[racer12306]

I think I may be using the wrong terminology. Let me explain the exact situation.


Part of our business involves inspecting manufacturing facilities. When we visit these facilities, sometimes they only have one product assigned to that facility. Other times there are two, three or ten products. A requirement on us is to review every product at least once every two years. I'm working on a process that will allow me to audit what our global inspectors our doing to make sure they are following this rule.

So where as normal sampling, such as to the MIL STD, requires 100$ inspection when the failure reaches a specific point, I'm not looking to do that. I just want to come up with a reasonable number of jobs every month to check. Then if there is a failure, I get in touch with that inspector and educate them on what they have been trained to do.

 

[wkcars]

Seems like what you're looking for is an auditing schedule? Something along the lines of risk based auditing? That's going to be product/supplier specific, and you would need to do some analysis to figure out the risks/priorities/frequencies etc. for your own organization.

 

[Turk]

Yea, I needed more information.

A reasonable number of jobs to check that will be statistically significant is 59 with zero failures.

If you use that number and something not good slips by, you'll have a statistically significant industry standard as a defensible position for you & your company.

 

[Shannow]

Weibull ?

Because the truth hurts sometimes.

 

[racer12306]

Thanks for the insight.

I've been thinking about this more, and I'm liking the number of 50 per month. The only specific data I want is who is following the rule. If I get a failure, that inspector is going to get a stern email and/or phone call and I will keep a closer eye on that inspector.

 

[Turk]

If you have 2-3 Inspectors, I would recommend you do a Gage R&R.

R&R is Repeatability & Reproducibility. You're identifying how repeatable each inspector is on a given number of defect types is, usually 10, and how reproducible it is between Inspectors.

If they do not pass, you'll then know the inspection procedures will need to be updated and the Inspectors re-trained and certified.

Look it up and design your experiment. It's straight forward & easy. Plus, you'll look like a star.