DMR: Delete or Merge Regressors for linear model selection

A backward selection procedure called delete or merge regressors (DMR) combines deleting continuous variables with merging levels of factors. The method assumes greedy search among linear models with set of constraints of two types: either a parameter for a continuous variable is set to zero or parameters corresponding to two levels of a factor are compared. DMR is a stepwise regression procedure, where in each step a new constraint is added according to ranking of the hypotheses based on squared t-statistics. As a result a nested family of linear models is obtained and the final decision is made according to minimization of the generalized information criterion (GIC, default BIC). The main function of the package is DMR, which is based on hierarchical clustering. Moreover, other functions for extensions of DMR method are given, such as stepDMR which is based on recalculation of t-statistics in each step and function DMR4glm for generalized linear models.

Downloads: