Semiparametric estimation of mean and variance functions for non-Gaussian data (Q2463671)

The author suggests an approach for modelling mean and variance functions semiparametrically that is appropriate for modelling continuous heterogeneous data and data exhibiting extra-binomial and extra-Poisson variation. He is interested in what factors affect the variability of the responses. That is why modelling mean and variance for possibly non-Gaussian data is important. The paper is concerned with estimating a mean and variance function in a very general way, modelling the responses using a double exponential family distribution with location and dispersion parameters described as an additive function of predictors. A simulation-based approach using adaptations of standard Markov chain Monte Carlo techniques is taken for carrying out the computations required for Bayesian inference in flexible double exponential regression models. For such models, an adaptive Metropolis algorithm of \textit{H. Haario, E. Saksman} and \textit{J. Tamminen} [Comput. Stat. 20, No. 2, 265--273 (2005; Zbl 1091.65009)] as a useful general approach to computation is presented. This approach is illustrated in several examples.