ABSTRACT
Although linear models with normal responses are very versatile in many applications, like any other models, they have limits. These models are not suitable when responses clearly follow a non-normal distribution, such as when they are discrete counts or binaryvalued responses, or they are strictly positive continuous-valued responses, or they exhibit substantial non-linearity in their relationship with the predictors. A motivation for using generalized linear models (GLIM's) is that they permit more general distributions than the normal for the response (Nelder and Wedderburn (1972) and McCullagh and Nelder (1989)). A GLIM generalizes a linear model by allowing the expectation of the response variable to be related to a parametric linear function of the predictors via a suitable link function. The usual normal linear models can be thought of as a particular type of GLIM's. This chapter describes GLIM's for non-normal responses.
