A0184
Title: FBMS: An R package for Flexible Bayesian Model Selection
Authors: Florian Frommlet - Medical University Vienna (Austria) [presenting]
Aliaksandr Hubin - NMBU (Norway)
Geir Olve Storvik - University of Oslo (Norway)
Jon Lachman - Stockholm University (Sweden)
Abstract: The R package FBMS has implemented a highly flexible approach to construct nonlinear parametric regression models. Bayesian inference is performed using a genetically modified mode jumping Markov chain Monte Carlo algorithm (GMJMCMC), which is at the same time used to hierarchically generate non-linear features. The flexibility of FBMS both pertains to the model itself as well as to the generation of non-linear predictors. With respect to the choice of models we will demonstrate how to apply FBMS for the generalized linear model and then go beyond to use pretty much any parametric model for which the likelihood function can be specified. With respect to the prediction function, the space of non-linear features which can be generated includes many familiar model families as special cases, like, for example, fractional polynomials, neural networks or logic regression. We will show how to make use of FBMS to perform variable selection for some specific families of non-linear features and illustrate the good performance of FBMS compared to some competitors.
