Downloaded from https://www.everyspec.com AMCP 70-18 has been determined for the particular tests. tests are "better" than nonpaiametric tests The nominal OC curve can be off by a factor since, tor a given amount of testing, more of 2 or so in a and j3, See Refs. 21 and 22 precise estimates are obtained from the for techniques of analyzing the seq-sample parametric tests, However, see Ref. l9 for plans, a more complete discussion of this point; , many- nonparametric tests are very "good", 3.6.3 SPECIAL TESTS and do not have the big disadvantage of having made the wrong distributional assump- When the "reject" decision is equivalent to tion.
"test the entire lot" rather than "destroy the lot" or something equally as drastic, the test Many specifications are written in terms of described in the following paragraph may parametric testing or in terms of such prop- have advantages for the producer, while re- ertie6 as mean-life which are not too suitable "taining protection for the consumer. for nonparametaic tests, For example, non- 4 parametric tests of central tendency apply to Use an rn-smple or seq-sample test, but the Cdf while the specification may be in move the "reject" part of the boundary to terms of mean life. This might require a "infinity", This means that a lot will never change in specifications.
by rejected; in lieu of that, one may have to test the entire lot before accepting it. But An incorrect assumption of the underlying that is what a "rejeot" decision would have failure distribution in a parametric test can meant, anyway. Details of such plans are lead to an OC curve that differs greatly from not readily available in the open literature, that planned, especially for small sample sizes. Ref. 26, sections B.0933 and B.0935 discut. This is not a problem in nonparametric tests. such plans, at least in principle.
Also, nonparametric tests are generally easy to conduct and evaluate, and often require An alternative is to put the "reject" line only counting, adding, subtractin3, or ranking; very high; so the probability of fully testing see Ref. 19. an acceptable lot is very small. 3-5.5 REPLACEMENT 3MA.4 ASSUMING A FAILURE LAW Replacement here is not used in its stath- In life tests, a failure distribution is almost tical sense of discrete nondestructive sampl- always assumed, and it usually 6b the constant ing (e.g., urn problems or decks of cards), failure-rate distribution.
The parameter to be judged is the failure rate (or, equivalently, Replacement tests are those In which failed its reciprocal), items are replaced by new items, so that the test stations are always full.
If the items are If one just considers the fraction good or complex, replacement may be interpreted to bad, then the parameter to be Judged is the be restoration of the failed item to new con- frection bad (or, equivalently, the fraction dition by repair or replacement of failed good), This assumption (the parameter components, In nonreplacement tests, failed "fraction bad" is the same for every item items arn not replaced or repaired; therefore, tested) is much more likely to be fulfilled the number of items on test decreases as life than is the assumption of a specific failure testing progresses.
distribution, "Generally, the effect of using a replace- Conventional wisdom states that parametric ment test is to decrease the calendar waiting .. •. •...,,, ....