Abstract
Consider the following experiment: At stress levels x1 < … < xk, n1, …, nk items are put on test. The failure times of those items failing before times t 1*, t k* are recorded, and the remaining failure times are censored. Let P(t, x) denote the probability that an item stressed at level x fails on or before time t. We consider models of the form P(t, x) = Φ[g(t, x)], where g is a real-valued function specified up to an unknown parameter c, and Φ is an unspecified continuous distribution function. Under these assumptions a Kolmogorov-Smirnov-type test is discussed as a method for estimating c as well as P(t o, xo) for specified t o and xo. The method is exact and does not require failures at unaccelerated stress levels. If n 1 = … = n k , the method extends to experiments that at each stress level are terminated after a specified number of failures is observed.