Relative efficiency of methods for estimating general gamma creep failure time distribution in presence of censored data

Little is known about the exact form of the creep failure time distributions, and the presence of unfailed tests is most data sets adds to the complexity involved in such identification. Monte Carlo techniques were used to assess the relative efficiency of various methods available for estimating the parameters of the generalised gamma failure time distribution. The censored likelihood procedure was shown to be the last biased and most efficient estimator, although sample size, the extent of censoring, and shape of distribution affected its relative efficiency. Ignoring unfailed test specimens always led to inefficient parameter estimates. Using the generalised gamma distribution, the censored likelihood technique was used to test the form of the failure time distribution for 9Cr-1 Mo steel tubes creep tested at 220 MN m(-2) and 500 degrees C. For this data set, about 25% of the test specimens were removed before failure had occurred. Results suggested that the failure time distribution is log normal. A wide variety of 3 parameter gamma distributions and the Weibull distribution were also shown to be consistent with these data. This is highly consistent with results from past studies using only uncensored data sets. The removal of the censored observations led to an il defined distribution. (C) 1996 The Institute of Materials.