Fitting Finite Mixtures of Linear Regression Models with Varying & Fixed Effects in R∗

A general model class of finite mixtures of linear regression models is presented. It allows (nested) varying and fixed effects for the regression coefficients and the variance. A combination of varying and fixed effects is useful in applications because it can be used to account for overdispersion as a nuisance parameter or to reduce the number of estimated parameters. In addition concomitant variable models for the component weights provide the possibility to partition the data into the mixture components through other variables. Maximum likelihood parameter estimation using the EM algorithm is outlined and the implementation in R by extending package flexmix is described. In this paper multinomial logit concomitant variable models are considered, but the provided infrastructure allows to easily define new concomitant models and rapid prototyping is possible if functionality already available in R can be used.