Defective subpopulation distributions, sometimes known as limited failure populations, arise when only a fraction of units are subject to some kind of an event. A flatting of points on a probability plot may imply a defective subpopulation. For subpopulations, it is important to test long enough to provide data that will be sufficient to decide if you have many units failing slowly or a small fraction failing rapidly.
In cases where there is a huge difference between the Wald and the Likelihood-ratio based confidence interval, choose the likelihood interval. View a video on this topic by Bill Meeker and a related JMP demo by Leo Wright.
This webcast is part of the Statistical Methods for Reliability with Dr. Meeker on-demand webcast series.