The Evaluation of a Concomitant Variable Behaviour in a Mixture of Regression Models

Abstract

Finite mixture of regression models are a popular technique for modelling the unobserved heterogeneity that occurs in the population. This method acquires parameters estimates by modelling a mixture conditional distribution of the response given explanatory variables. Since this optimization problem appears to be too computationally demanding, the expectation-maximization (EM) algorithm, an iterative algorithm for computing maximum likelihood estimates from incomplete data, is used in practice. In order to specify different components with higher accuracy and to improve regression parameter estimates and predictions the use of concomitant variables has been proposed. Based on a simulation study, performance and obvious advantages of concomitant variables are presented. A practical choice of appropriate concomitant variable and the effect of predictors' domains on the estimation are discussed as well.

Keywords

• Mixture of regression models
• linear regression
• EM algorithm
• concomitant variable