Yah, sample size and AQL are important to know here. You need to basically calculate the probability of failure in your population given the zero probability rate in your samples.
For a first stab at it, try a binomial calculator (RAC has a good one on the web, which they moved and I can't find, but there are others out there -- and a pretty good article http://src.alionscience.com/pdf/OC_Curves.pdf) and feed in the number of failures (0) and the number of trials (your total number of samples) to find out what confidence you can claim on your desired defect rate. For instance, 29 good out of 29 trials only buys you about 95% confidence that 90% of an infinite population is good. Put another way, if you can stand as much as 10% bad parts, you can be 95% sure you met your goal.
You probably already well know that you need a boatload of total samples to confirm goodness in pass/ fail data. Playing with the calculator will reaffirm that! So many samples that it's quite possible that your process changes significantly during the time it takes to draw the samples. The samples you drew at the beginning of the process may not belong grouped with those taken recently.
You might also want to look into Bayesian techniques for characterizing low failure rates. Ugly math, but works pretty slick.
Community metadata Could not be loaded. No pre-configured community forum parent found for the incoming fid: . You need to add this forum into custom_code/community.sql
Send to friend
Comments
11/22/2005
What subgroup sample size have you been using and what is considered an acceptable quality level for the product you're checking?
12/8/2005
Yah, sample size and AQL are important to know here. You need to basically calculate the probability of failure in your population given the zero probability rate in your samples.
For a first stab at it, try a binomial calculator (RAC has a good one on the web, which they moved and I can't find, but there are others out there -- and a pretty good article http://src.alionscience.com/pdf/OC_Curves.pdf) and feed in the number of failures (0) and the number of trials (your total number of samples) to find out what confidence you can claim on your desired defect rate. For instance, 29 good out of 29 trials only buys you about 95% confidence that 90% of an infinite population is good. Put another way, if you can stand as much as 10% bad parts, you can be 95% sure you met your goal.
You probably already well know that you need a boatload of total samples to confirm goodness in pass/ fail data. Playing with the calculator will reaffirm that! So many samples that it's quite possible that your process changes significantly during the time it takes to draw the samples. The samples you drew at the beginning of the process may not belong grouped with those taken recently.
You might also want to look into Bayesian techniques for characterizing low failure rates. Ugly math, but works pretty slick.
Ready to convert to variables charts yet? -- KLJ