A comparison of two sequential Metropolis-Hastings algorithms with standard simulation techniques in Bayesian inference in reliability models involving a generalized gamma distribution (Q5936987)

In some previous papers [e.g., Sand. J. Stat. 25, No. 2, 271-292 (1998; Zbl 0910.62094)], the first two authors introduced a generalized gamma distribution for which: (a) A Bayesian inference in exponential survival models with left censoring is naturally derived. (b) The weighted sum of products of the generalized gamma distribution is proved to be a conjugate prior for the parameters of component lifetimes, having autopsy data in a Marshall-Olkin shock model. (c) A similar result as in (b) is shown for independent, exponentially distributed component lifetimes in a model with partial monitoring of components, having applications to preventive system maintenance. The aim of the present paper is to investigate the computational aspects of some models in reliability, based on the results (a)--(b)--(c). Two new types of sequential Metropolis-Hastings algorithms are defined, such that to be used in Bayesian inference when exact methods are intractable. The exact analytical solutions are available as a basis of comparison, and a quality criterion is used to show the ability of the proposed algorithms to reproduce marginal posterior densities close to the true curves in reasonable computation time. Among the Metropolis-Hastings algorithms based on sampling from a fixed distribution, sampling importance resampling (SIR) algorithm seems to be the fastest and most flexible. Rejection sampling (RS) algorithm has a limited applicability, whereas parametric independent chain (PIC) Metropolis-Hastings algorithm works well if the prior distribution is in reasonable accordance with the data. Even SIR algorithm is shown to have trouble when large discrepancies occur between the prior assessments and the data. The new parametric, sequential Metropolis-Hastings PAIC (parametric adaptive independent chain), PSAIC, PSIC algorithms, generalizing PIC ones, are proved to be superior in coping with such discrepancies and, any way, to work substantially better than the PDC (parametric dependent chain) algorithms. The idea behind the proposed parametric, sequential algorithms seems to be quite general and potentially useful in various other parametric classes of distributions, such as convex combinations of products of (generalized) gamma distributions, normal distributions, combinations of uniform distributions, etc.