SuddenDeath Test

[Fudjie]

I am very interested in the Sudden-Death testing discussed in the July edition of HotWire. My question is, using Weibull++, is there any difference in the result (& confidence levels/accuracy) between simple suspension appraoch and the sudden-death approach (in groups)? Referring to the example in the HotWire, would testing ALL the units, at the same time, to an end-time of 185 be the same as testing the units in groups up to an end-time of 300?
Thanks very much.
Fudjie

[RS Support]

What the example points out (see http://weibull.com/hotwire/issue5/hottopics5.htm) is that if one tested 8 groups of 5 units until the first failure in each group, or if one tested 40 units until the 8th failure, the expected test duration would be the same. In this case to get to the 8th failure one would have to run it until 300 hrs.
The purpose of the example is to illustrate that the analysis can be easily done in Weibull++, simply by assigning censoring times to the surviving units, and not to recommend one test over the other. The advantage of ST testing is that if the facility can only test 5 units at a time, then ST is preferred. The disadvantage is that the test is sequential and would take much longer to complete. One other advantage of running all units at the same time is that the remaining units will realize higher censoring times, since the test will not be terminated until the last failure.

[greg peine]

If a second group of 40 units from a differnt supplier were tested in 8 groups of 5, how does one determine the difference between the groups using confidence? In particular what approach should be used to determine the confidence in the difference at the B50 between the lower confidence bound of the younger group and the upper confidence bound of the older group?
Thanks,
Greg

[RS Support]

The question asked could be more generalized on how do you compare two sets:

One could compare lifetimes – and the probability that lifetimes from A are greater than B, also one could compare to see if they are significantly different using contour plots etc. (see discussion at http://www.weibull.com/LifeDataWeb/tests_of_comparison.htm), or even look at the data sets with confidence bounds for overlap…

What you are asking (if I understand the question) is how do you say that the result at given percentage point is X times better than another result at a given confidence level.

One approach is to use concepts from accelerated life testing (using a binary response variable for each set i.e. 0 or 1 – see http://www.weibull.com/acceltestwebcontents.htm ), determine an acceleration factor (really a better than factor in this case) with confidence intervals.

Another approach is to get a function that provides a ratio of the two B50 values and find the confidence bounds on that function at a given confidence level -- if using Fisher Matrix expand (see http://www.weibull.com/LifeDataWeb/approximate_estimates_of_the_mean_and_variance_of_ a_function.htm)