createGroupset {ecpc}

Create a group set (groups) of variables

Description

Create a group set (groups) of variables for categorical co-data (factor, character or boolean input), or for continuous co-data (numeric). Continuous co-data is discretised in non-overlapping groups.

Usage

createGroupset(values,index=NULL,grsize=NULL,ngroup=10,
                decreasing=TRUE,uniform=FALSE,minGroupSize = 50)

Arguments

values

Factor, character or boolean vector for categorical co-data, or numeric vector for continuous co-data values.

index

Index of the covariates corresponding to the values supplied. Useful if part of the co-data is missing/seperated and only the non-missing/remaining part should be discretised.

grsize

Numeric. Size of the groups. Only relevant when values is a numeric vector and uniform=TRUE.

ngroup

Numeric. Number of the groups to create. Only relevant when values is a numeric vector and grsize is NOT specified.

decreasing

Boolean. If TRUE then values is sorted in decreasing order.

uniform

Boolean. If TRUE the group sizes are as equal as possible.

minGroupSize

Numeric. Minimum group size. Only relevant when values is a numeric vector and uniform=FALSE.

Details

This function is derived from CreatePartition from the GRridge-package, available on Bioconductor. Note that the function name and some variable names have been adapted to match terminology used in other functions in the ecpc-package.

A convenience function to create group sets of variables from external information that is stored in values. If values is a factor then the levels of the factor define the groups. If values is a character vector then the unique names in the character vector define the groups. If values is a Boolean vector then the group set consists of two groups for True and False. If values is a numeric vector, then groups contain the variables corresponding to grsize consecutive values of values. Alternatively, the group size is determined automatically from ngroup. If uniform=FALSE, a group with rank $r$ is of approximate size mingr*(r^f), where f>1 is determined such that the total number of groups equals ngroup. Such unequal group sizes enable the use of fewer groups (and hence faster computations) while still maintaining a good ‘resolution’ for the extreme values in values. About decreasing: if smaller values mean ‘less relevant’ (e.g. test statistics, absolute regression coefficients) use decreasing=TRUE, else use decreasing=FALSE, e.g. for p-values. If index is defined, then the group set will use these variable indices corresponding to the values. Useful if the group set should be made for a subset of all variables.

Value

A list with elements that contain the indices of the variables belonging to each of the groups.

Author(s)

Mark A. van de Wiel

See Also

Instead of discretising continuous co-data in a a fixed number of groups, they may be discretised adaptively to learn a discretisation that fits the data well, see: splitMedian.

Examples

#SOME EXAMPLES ON SMALL NR OF VARIABLES

#EXAMPLE 1: group set based on known gene signature (boolean vector)
genset <- sapply(1:100,function(x) paste("Gene",x))
signature <- sapply(seq(1,100,by=2),function(x) paste("Gene",x))
SignatureGroupset <- createGroupset(genset%in%signature) #boolean vector

#EXAMPLE 2: group set based on factor variable
Genetype <- factor(sapply(rep(1:4,25),function(x) paste("Type",x)))
TypeGroupset <- createGroupset(Genetype)

#EXAMPLE 3: group set based on continuous variable, e.g. p-value
pvals <- rbeta(100,1,4)

#Creating a group set of 10 equally-sized groups, corresponding to increasing p-values.
PvGroupset <- createGroupset(pvals, decreasing=FALSE,uniform=TRUE,ngroup=10)

#Alternatively, create a group set of 5 unequally-sized groups,
#with minimal size at least 10. Group size
#increases with less relevant p-values.
# Recommended when nr of variables is large.
PvGroupset2 <- createGroupset(pvals, decreasing=FALSE,uniform=FALSE,
                              ngroup=5,minGroupSize=10)

#EXAMPLE 4: group set based on subset of variables,
#e.g. p-values only available for 50 genes. 
genset <- sapply(1:100,function(x) paste("Gene",x))
subsetgenes <- sort(sapply(sample(1:100,50),function(x) paste("Gene",x)))
index <- which(genset%in%subsetgenes)

pvals50 <- rbeta(50,1,6)

#Returns the group set for the subset based on the indices of 
#the variables in entire genset. 

PvGroupsetSubset <- createGroupset(pvals50, index=index,
                                   decreasing=FALSE,uniform=TRUE, ngroup=5)
#append list with group containing the covariate indices for missing p-values
PvGroupsetSubset <- c(PvGroupsetSubset,
                      list("missing"=which(!(genset%in%subsetgenes))))

#EXAMPLE 5: COMBINING GROUP SETS

#Combines group sets into one list with named components. 
#This can be used as input for the ecpc() function.

GroupsetsAll <- list(signature=SignatureGroupset, type = TypeGroupset,
                     pval = PvGroupset, pvalsubset=PvGroupsetSubset)
               
#NOTE: if one aims to use one group set only, then this should also be
# provided in a list as input for the ecpc() function.

GroupsetsOne <- list(signature=SignatureGroupset)