Abstract:

In this thesis, we present solutions for the controlled calibration problem under the Bayesian perspective to statistical inference. The first problem considered is related to the linear case with elliptically distributed errors. For the elliptical dependent model it is shown that the posterior distribution for the quantity of interest coincides with the posterior distribution obtained under the normality assumption with an improper prior distribution. A conjugate prior analysis is also considered. The coincidence of the inference is not noted for the independent elliptical model. As consequences of some prior specifications and representability of the elliptical model, general expressions are obtained for the posterior distributions, which are characterized as mixtures of known distributions. Moreover, simple conditional posterior distributions of one parameter given the others were obtained which allowed the use of the Gibbs sampler to obtain approximations to the posterior distributions of the quantities of interest. Subsequently, the calibration problem for nonlinear situations was considered under the assumption that the response variable is categorized. A generalization of the well know probity model was considered, by taking an elliptical distribution function as the link function. For this case, an asymptotic approximation was obtained for the posterior distribution. A MCMC (Monte Carlo Markov Chain) solution is considered for the binomial model. For the multinomial model, a MCMC solution is present and the conditional distributions required for implementing the approach are obtained.