Practical Methods using Sampling by Variables

[medical_eng]

Hi everyone,

Although not an expert, I am familiar with Z1.4 and I see that people generally have an easy time understanding it (QA staff up to the General Manager, suppliers, customers, etc.). Rejecting a lot is usually not taken lightly and if a lot is to be rejected it’s easy to point to the rejects/nonconformances to substantiate the decision.

One approach is to convert any variables type measurement to a pass/fail comparison and use Z1.4. This may be common practice in many companies but I don’t know. Supposedly, though, if you sample by variables you can reduce the sample size greatly, saving valuable time and cost.

I have two questions regarding sampling by variables:

1) I see three methods available and can’t decide which one would work out better in practice:
a. Using Z1.9 (which I don’t know very well yet).
b. Using ISO 16269-6 or similar (using the ‘k’ multiplier) and checking that the calculated intervals remain within the specification interval.
c. Calculating the capability indices and comparing that to a target (1.33, 1.67 or whatever).
Does anyone have any practical experience on this?

2) The whole procedure of sampling by variables and then comparing the calculated results to a target appears to involve ‘extrapolation’. Thus one could be in a position where all the sample results show a pass, but the calculations indicate that the lot should be rejected. How does one explain the need to throw out (or rework/sort/etc.) a lot when no actual bad parts have been found? Have any of you had this situation and what was done?

Thanks!

 

[DRAMMAN]

You are zeroing in on the actual value of a variables sampling plan. It is the same reason why try to use variable data to evaluate the quality of something. Both sampling approaches are trying to determine the same thing...does my lot of material have an acceptable defect percentage. People falsley believe that not finding defects in my sample indicates there are zero defects in the lot. All it means is that you have a confidence level (typically 95%) that the defect percentage is less than the AQL level. It also means ther eis a 5% risk the defect level is greater than the AQL level. Think about how many times production finds a high defect percentage even though no defects were found durring the attribute acceptrance sampling. That same shipment may have been rejected had variable sampling been used.

Those shipments with no sampling defects that are rejected with variable sampling will likely have defects in the lot. If you screen it 100% you will probably find defects. It is also possible you will not find any if the supplier somehow inspected them out. The valuable bit of information is you know know how the supplier's process is operating. Those defects are likely increasing your costs since your are most likely somehow paying for them in your price. Know you can go back to the supplier and push for process improvements using your capability data.

 

[medical_eng]

No other comments or thoughts from others with experience in this area?

Hopefully someone has success with sampling by variables and can share their thoughts on the preferred methodology from the three suggested.

Thanks for your reply DRAMMAN, and while I don't disagree with your comments it still doesn't get to the root of the other half of my question. Perhaps stated another way, I need to convince the skeptics (myself included) that the system can 'work'. Imagine oneself with a batch of parts and the calculated result (based on a normality assumption) doesn't meet the target requirement/goal. So, in a contrived hypothetical situation, the general manager (who is not a guru in statistics) is now involved and asks "Show me the bad parts...". Under pressure of the situation I can see the conversation going bad really fast, which would be a mistake and a shame. What have others done in this situation or how have they avoided it to begin with?

One potential solution that I see is to pull the sample quantity by attributes to the required AQL but test a lesser quantity by variables. If everything works out you have saved labour, a good thing for everyone. If things don't work out on the variables indices, you keep going for the full sample size. If this finds no defects then the lot is deemed good and the variables indices are given no standing (maybe the standard deviation was a little off, throwing the numbers - opportunities for future improvement). If you fail both checks then you will have shown due diligence with the extra step and you will have 'defects' in hand to substantiate the decision. Any opinions on this proposal?

Thanks!