Inference procedures for the linear failure rate model

Maximum likelihood and least-squares-type estimation of the linear failure rate (LFR) distribution are studied for type II censored samples. It is shown that a particular structural property of the LFR greatly facilitates application of the EM algorithm for computing the MLE's. Also, an apparently ad hoc method, which rests on maximization of a pseudolikelihood, is shown to produce the MLE's. A finite-sample exact confidence procedure is developed as an alternative to the existing MLE-based large-sample procedure that suffers from poor coverage probabilities even in moderately large samples. Also, asymptotic efficiencies of the least-squares-type estimates relative to the MLE's are derived. Numerical computations show that these can be quite low in many cases.