No label defined (Q4529786)

The text under review covers the relatively new field of nonlinear Bayesian modeling. The ideas of nonlinear modeling and Bayesian theory are brought together to provide methods for producing probabilistically sound inferences. It is emphasized that the main difficulties lie in the implementation of the methods, as well as in understanding which model could be expected to perform well in what situation. The authors discuss the strengths and weaknesses of the introduced models. The book consists of nine chapters, two appendices, bibliography, and subject and author indices. Chapter 1, ``Introduction'', states the problem of quantifying the relationship between a response variable of interest and measurements taken on a set of possibly related observations. Since the true functional form between the input and output variables is usually unknown, the authors discuss questions that influence how they proceed: (i) What type of approximating functions should be used? (ii) How does one know that the `best' approximation to the truth has been found? (iii) How can one incorporate any available quantitative or qualitative knowledge about the relationship? Then an outline of the book is provided. Chapter 2, ``Bayesian Modelling'', begins by giving a rationale of why a Bayesian approach to predictive inference was chosen. It then goes on to demonstrate how such an approach can be implemented. This is described with reference to the regression problem, which dominates the first chapters. Chapter 3, ``Curve Fitting'', goes through the process of fitting simple Bayesian models, both analytically and via posterior simulations. The curve fitting model is used to demonstrate simple applications of the methods with a view to aiding the understanding of readers new to the field. Chapter 4, ``Surface Fitting'', then introduces models for fitting regression surfaces to data. It makes a distinction between models that are more suited to problems when there are many predictors and those that work well in fewer dimensions. In Chapter 5, ``Classification Using Generalised Nonlinear Models'', the authors show how to make posterior inference in those more difficult problems when analytic simplification is not available and simulation methods for high-dimensional posteriors are required. It is illustrated that generalized linear models provide a powerful toolkit for modeling in a variety of situations when the errors in the observations are not additive and normally distributed. Chapter 6, ``Bayesian Tree Models'', covers Bayesian extensions to tree modeling. These are mainly used for the classification problem, although the authors include discussion of regression trees. Chapter 7, ``Partition Models'', extends the general method to partition models. These are similar to trees but allow for more flexible partitioning of the predictor space into regions where data are assumed homogeneous. Chapter 8, ``Nearest-Neighbour Models'', goes on to describe Bayesian nearest-neighbor modeling. Again, this method uses local structures in the data to provide a good way of making predictions. Finally, Chapter 9, ``Multiple Response Models'', covers multiple response modeling when inference is made about a vector of response variables given the predictors. This generalizes the single response case, studied throughout the rest of the book.