Downloaded from https://www.everyspec.com *AR AM 701-197 -where S8 s the sum of squares of deviations for the sample(s) would be exceeded 40 per- about the sample mean, and N is the number cent of the time, then it is not likely that one of items in the sample-is an s-unbiaed esti- method is better than another. The percen- mator of 02 (the true value of the variance), tWe chosen (0.1 percent, 40 percent, etc.) is but S is an s-biased estimator of o. (The called the s-significance level. In practice, en- square root function is not linear.) Another gineers want the effect to be s-significant at a example is 1/X, the reciprocal parameter for 20 percent level or less.
an exponential distribution. An s-unbiased w Regardless of the outcome of the statisti- timator for 1I/ is the sample m.an, but the cal test, the engineer wants the effect to be of reciprocal of that estimator is an s-based esti. engneering importance. It is possible to take mator of X. (The reciprocal is not a linear a sample small enough so that no matter what function.) the actual difference is, it will not be s-signifi- How is an engineer to know what func- cant because the uncertainties due to too few tion of the parameter ought to be s-unbiased? data overwhelm all other considerations. On He doesn't.
In general, reliability engineers the other hand, it is also possible to take so can ignore a-bias of estimators; they need onqj much data that the difference will be s-eignifi- be concerned about a-efficiency. cant, no matter how small the effect. Tests of s-significance suffer from being equivalent to 4-2.4 UNCERTAINTY point estimates. Engineers would rather esti- mate the difference between two methods Any estimates of parameters ought to be and the it icertainty in that estimate. This pro- accompanied by an estinate of the uncertain- cedure is discussed in par. 4-4 on s-confidence ty involved. Two common methods of indi- statements.
cating uncertainty are the covarance matrix and s-confidence intervals. The reliability en- 4-4 s-CONFIDENCE STATEMENTS gineer need not know how to get them, only how to use them. As with s-significance there is an impor- tant difference between the engineering and 4-3 TESTS OF s-SIGNIFICANCE statistical concepts. s-Confidence i. a statisti- cal concept with a very special, exact mean- The most important thing about s-signif- ing. Don't use the concept without under- icance is what it isn't; it is not "engineering standing that meaning. importance". s-Significance is concerned with An example statement is a good way to tests t&at are run to see if one thing is differ- understand the concept.
ent from another. A statistical model is for. mulated and measurements (test.) are made. "The true improvement in fatigue on the sample(s) to meastvie the difference in strength (method B over method A) lies the items of the sampl. . F r example, does between -1.7 and + 10.9 kips/in. s at a heat-treating method A prodent 'uetter fatigue 90 percent s-confidence level." properties than heat-treating nietu'oa b., Usu- The 90 percent s-confidvnc' level means that ally the statistical hypothesis is made that 90 pment of the times that our' goes through there is no difference. Then the statistical dis- the statistical manipulations as doune for this tribution of the test statistic is calculated.
In example, the resulting statement will be cor- the example, the test statistic might be the rect; 10 percent of the time it will be wrong. difference in average fatigue-strengths at 107 The -1.7 and + 10.9 kips/in.2 are called the cycles of stress. The value of that test statistic s-confidence limits. for the sample(s) is measured and compared For a given set of sample measurements, with the distribution. If a value as large as the higher the s-confidence level is, the wider observed would occur only 0.1 percent of the time or less, the effect (difference) is not like- the s-confidet,,r limits will be.
ly to have been a chance observation, but is An engineer might look at the s-confi- likely to be due to one method being better dence statement and say, "Even if the im- than another. If the value of the test statistic provement in fatigue strength were as good as 4-2