Decomposition estimation: A generalized linear model based technique for fitting multi-linear regression models to grouped data in GLIM 4

Fitting regression models with GLIM is being approached in many different ways if non-linear models are outside the class of GLMs GLIM has been intended for. We propose a technique which is based on: (1) forcing a multinomial (here binomial) sampling scheme by grouping the random variable of interest (e.g. failure time), and (2) separating or ‘decomposing’ covariate effects by repeating the data, each time with an appropriate link to some systematic linear structure. This paper focuses on our GLIM4 implementation of the technique: GLIM DECO, which is a menu-driven collection of macros, developed to avoid awkward programming for the user. Examples demonstrate the wide range of applicability of our technique.