Sample size for capability

[cjholl0]

I work for an automotive supplier and our QA dept routinely uses a sample size of 32 for CpK data. I've asked several people what the basis for this number is but nobody has been able to tell me, just that's what we've always done. It seems to me that there would have to be some consideration of the qty of parts in question. If we machine say 100,000 aluminum bodies in a month, would taking 32 pieces from one of the machines really tell you what the capability for that machine to hit that dimension is? Could anyone shed some light on this for me? Anybody know where 32 comes from? More generally, how can you choose a sample size? If, say, you have a sample of 60,000 pcs that may have some defect, how can you pick a qty to measure to gage the number of defects and how large (dimensional variation) the defect may encompass? Thanks in advance for your thoughts.

 

[dingy2]

The sample size must be 30 piece or over for the data to form a normal distribution.

Here is how I would chose a sample size.

Look at your process. If it is a machine that produces 1 product per cycle, then I would probably use 50 consecutive pieces. I know that it has to be over 30 so one could drop it down to that value.

If you process has 3 stations affecting the products, I would not use a sample size of 30 per station (as a lot of suppliers do) but I would make sure that the sample size (probably 60 - 20 per station) is homogenous.

If we had a 12 station machine can you imagine the cost of running 30 pieces per station. Again use a homogenous sample size.

Review the histogram making sure that the data reflects a normal distribution. If 1 station is out to lunch relative to the other stations, one could either change the tooling to bring it more in line with the other stations OR run 30 pieces from that station only.

The Ppk (not Cpk) value in this situation would be the worst one if you did run 1 station separately.

Hope this helps.



Dave D.

http://www.qmsi.ca

 

[cjholl0]

What exactly is meant by homogeneous? For the 3 station example you mentioned not running 30/station, instead 20/station. Is the key to run the same number per station?

Also, why is 30 needed for a normal distribution? If you have only 20 do certain statistical tools not apply? Does it have something to do with confidence interval?

 

[dingy2]

Yes - run the same number per station.

The 30 parts minimum is the number from any stats book. When you calculate the upper and lower control limits using this number, one achieve a 99.73 % confidence level. Anything under this magic number of 30 would end up with less confidence.

The confidence interval is the difference between the upper and lower control limit. The confidence level used in industry today is 99.73% or +/- 3 estimated standard deviations.

Dave D.

http://www.qmsi.ca