Re: AQL for Zero Defects
Acceptance (random) sampling from a completed lot is not the same as SPC where a non-homogeneous process stream is sampled in a stratified manner in time sequence. The statistics are different (and not just theoretically but in reality). Random sampling from a non homogeneous lot actually 'homogenizes' the distribution of results within the sample. AND for categorical sampling the underlying distribution is irrelevant as the number of defects in the sample is a function of probability theory not distributional statistics. The 'normality' of the distribution - in this case - is irrelevant.
Point well taken. It is good to bring up what the
statistical requirement of 'random" really means, and if you want your sampling to work, you must
ensure that level randomization - especially the smaller the number of resident defects in the population. Just grabbing a handful of parts -
assuming the process had randomized them - will not meet that requirement if the population is indeed stratified. It is clear that true random sampling will 'homogenize" the lot. For example, if lot sampling was from a rotating drum of bingo balls - and you pulled enough of them - you would be well represented of the probability of possible defects in the lot. That is the beauty of a perfect statistical example. If you can sample
that randomly - yes, you will be in great shape. However, the reality sampling with such a clean randomization in an industrial setting (especially incoming receiving) is not always met - due to packaging considerations, etc. Anything less than perfect randomization
can create sampling error causing erroneous conclusions.
So, the incoming distribution will not matter with sampling under those ideal conditions, and - conversely - you
cannot determine the incoming distribution from sampling for similar reasons (those of you trying to calculate Cpks from incoming receiving sampling...
not a good idea).