Karen,
There seem to be several issues here that are tied together. The ultimate concern is that you are getting complaints about quality, which will affect customer satifaction and eventually profits. So something should change! One question is whether even your best parts that past the test are good enough for the customers need.
Assuming that you make at least some good parts that work for the customer, the direct problem for you is how to test parts to sort the good from the bad. It appears that when you test multiple times, you get different results. This could be due to I) the part actually is borderline defective and the test has a difficult time judging or II) the test itself is defective (a large value of alpha and/or beta). Do you have a feel for whether either or both of these is occuring?
For case I, multiple testing will simply push the borderline parts into the "accept" pile.
How to test depend on just what kind of expectations you have. For example, suppost you have reason to believe that only 1 out of 1000 parts is bad, but that the test will reject a good part 1 out of 5 times and always rejects bad parts (i.e. a Type I error with a large alpha). If you test 1000 parts, you would expect it to reject the one bad part, but also to reject (on average) 200 good parts. If you retest these rejected parts, then you will again reject the one bad part, but now reject just 40 good parts. One more pass and you are down to 9 rejects on average (the one bad one and 8 good ones).
But we can revese the situation as well. Suppose that 100 out of 1000 are bad (i.e. p = 0.1), but the test accepts 1 out 5 bad parts and accepts all the good parts (i.e. a Type II error with a large beta). After the first pass, you accept the 900 good parts, but also 20 of the 100 bad parts. If you retest the 80 rejects, you will accept another 16. The third pass will accept another 13 bad parts.
In the first case, each extra pass increases adds more good parts to the "accept" pile. In the second case, each extra pass adds extra bad parts to the "accept" pile. Changing the numbers will change the outcomes along this continuum. The "best" testing plan will depend on several factors - alpha, beta, p, cost of testing, the cost of scrapping a part, the value of selling a good part, and cost of selling a bad part come to mind.
(I can think of one other variation - the process of going through the testing could improve the product and actually turn it from bad to good. Perhaps some resin has to set and just the extra time from the first test to the second gives it time to cure. Perhaps the technicians wiggle the parts or brush off bits of dust to improve the performance.)
Sorry - I seem to have written quite a bit but I don't have a definite answer. Bottom line - most of the time I would tend to vote against repeated testing. Two cases where it could be appropriate would be
1) the parts are valuable and alpha is large and you retest the rejects (you are scrapping a lot of expensive good parts).
2) the cost of shipping bad parts is high and the beta value is high and you retest the accepted parts (you are selling a lot of bad parts and you want to make sure the parts you accept really are good).
Tim F
P.S. Has anyone seen an analysis like this? I might be valuable to have a formula to decide when multiple testing is worthwhile and how many times to repeat the test. I think I could come up with something of the sort if I got motivated.