ABSTRACT
This chapter discusses the Alternating Conditional Expectation (ACE) algorithm, a nonparametric generalization of the additive model that fits an additive model as part of an alternating estimation procedure. It explains the two approaches for estimating transformations that are designed specifically for regression problems. The first approach generalizes the well known Box-Cox procedure, while the second aims at variance stabilization. The papers by Bickel and Doksum Box and Cox, and Hinkley and Runger contain a lively series of arguments about inference for the model. The estimation of a response transformation makes it more difficult to make inferences from the fitted model about the predictors. Like the linear-regression model, it is useful when the assumed model is appropriate, but may be ineffective when it is inappropriate. The Additivity and Variance Stabilization procedure seeks transformations that achieve additivity and a homogeneous variance, and is more directed towards regression problems than is ACE.
