Delete or Merge Regressors for linear model selection.
Description
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.