obtainHierarchy {ecpc}

Obtain hierarchy

Description

This function obtains the group set on group level that defines the hierarchy; if a group of covariates g is a subset of group h, then group h is an ancestor of group g (higher up in the hierarchy). This hierarchy is used in adaptively discretising continuous co-data.

Usage

obtainHierarchy(groupset, penalty = "LOG")

Arguments

groupset

Group set of groups of covariates with nested groups.

penalty

Default: "LOG" for a latent overlapping group approach (currently the only option in ecpc)

Details

We use the latent overlapping group (LOG) lasso penalty to define the hierarchical constraints as described in (Yan, Bien et al. 2007); for each group g of covariates, we make a group on group level with group number g and the group numbers of its ancestors in the hierarchical tree. This way, group g can be selected if and only if all its ancestors are selected. This function assumes that if group g is a subset of group h, then group h is an ancestor of group g. Note that this assumption does not necessarily hold for all hierarchies. The group set on group level should then be coded manually.

Value

A group set on group level defining the hierarchy.

References

Yan, X., Bien, J. et al. (2017). Hierarchical sparse modeling: A choice of two group lasso formulations. Statistical Science 32 531-560.

See Also

splitMedian to obtain a group set of nested groups for continuous co-data.

Examples

cont.codata <- seq(0,1,length.out=20) #continuous co-data
#only split at lower continous co-data group
groupset <- splitMedian(values=cont.codata,split="lower",minGroupSize=5) 
#obtain groups on group level defining the hierarchy
groupset.grouplvl <- obtainHierarchy(groupset)