Chapter 30: The GAM Procedure (Experimental)

Overview

The GAM procedure fits generalized additive models as those models are defined by Hastie and Tibshirani (1990). This procedure provides an array of powerful tools for data analysis, based on nonparametric regression and smoothing techniques.

Nonparametric regression relaxes the usual assumption of linearity and enables you to uncover structure in the relationship between the independent variables and the dependent variable that might otherwise be missed. The SAS System provides many procedures for nonparametric regression, such as the LOESS procedure for local regression and the TPSPLINE procedure for thin-plate smoothing splines. The generalized additive models fit by the GAM procedure combine

• an additivity assumption (Stone 1985) that enables relatively many nonparametric relationships to be explored simultaneously with

• the distributional flexibility of generalized linear models (Nelder 1972)

Thus, you can use the GAM procedure when you have multiple independent variables whose effect you want to model nonparametrically, or when the dependent variable is not normally distributed. See the Nonparametric Regression section on page 1569 for more details on the form of generalized additive models.

The GAM procedure

• provides nonparametric estimates for additive models

• supports the use of multidimensional data

• supports multiple SCORE statements

• fits both generalized semiparametric additive models and generalized additive models

• enables you to choose a particular model by specifying the model degrees of freedom or smoothing parameter

Experimental graphics are now available with the GAM procedure. For more information, see the ODS Graphics section on page 1581.