CMStatistics 2018: Start Registration
View Submission - CMStatistics
Title: A fast estimation method in nonparametric additive location-scale model based on Bayesian P-splines Authors:  Philippe Lambert - Universite de Liege / Universite catholique de Louvain (Belgium) [presenting]
Abstract: In a previous publication on nonparametric additive location-scale models for interval censored data, it has been explained how Bayesian P-splines could be used in regression models to specify a smooth error density and the joint (possibly) nonlinear effects of covariates on location and dispersion. That methodology extends traditional additive regression models by releasing the parametric constraint on the error distribution and by acknowledging that covariates can affect multiple aspects of the conditional distribution in a non trivial way. These extensions are very attractive and practically useful, but have an important computational cost following from the use of the Metropolis-within-Gibbs algorithm in a richly parameterized model. We show how Laplace based approximations to the marginal posterior distributions of smoothness parameters can be used to set up a quickly converging iterative algorithm to select penalty parameters and to estimate the spline parameters in the pivotal distribution and in the additive components for location and dispersion in a fast and reliable way, as confirmed by simulation results. We conclude the presentation with an application to survey data.