R/modelMatrix.R

In dhelkey/dgrank: Construct Draper-Gittoes Rankings of Institutions

modelMatrix = function(x,
	interactions = FALSE, 
	sparse = TRUE,
	intercept = FALSE){
	#' Create design matrix from data frame
	#'
	#' \code{modelMatrix} takes in a data.frame and encodes it as a model.matrix object.
	#' Factors use treatment ("one-hot") encoding, creating indicator variables from categorical
	#' and binary variables.
    #' Note that names of categorical variables will change slightly 
	#'	(n-1 unique values appended to name)
	#'
	#' @param x  A data.frame with finite, non-missing elements.
	#' @param interactions Logical; if \code{TRUE} then all interactions included.
	#' @param sparse Logical; if \code{TRUE} then output a sparse Matrix object.
	#' @param intercept Logcal; if \code{TRUE} then include a leading column of 1's.
	if(is.null(x))return(NULL) #Dealing w/ empty data
	if (sum(!sapply(x, is.finite)) != 0){
		stop('Creating model matrix requires all input elements be finite')}
	
	#Set R to use 1-hot encoding
	options(contrasts = rep("contr.treatment", 2))
    model_formula = as.formula( ~.) #All first order Effects
    if (interactions){model_formula = as.formula( ~(.)^2)} #1st order and interactions	 
   mat = model.matrix(model_formula, data = data.frame(x))
	
    if (!intercept){
		#Remove intercept column, making sure names are preserved
		n_vec = colnames(mat)
		mat = data.frame(mat[ ,-1])
		colnames(mat) = n_vec[-1]
	}
    #Drop any empty levels
	n_vec = colnames(mat)
	nonempty_indices = apply(abs(mat),2, sum) != 0
	mat = data.frame(mat[ ,nonempty_indices])
	colnames(mat) = n_vec[nonempty_indices]

	mat = as.matrix(mat)
	if (sparse){mat = as(mat,'sparseMatrix')}
   return( mat)
}