R/method_predict.R

In BMSC: Bayesian Model Selection under Constraints

#' Compute predictions from constraint estimation model
#' @description Computes prediction from model of class
#' \code{\link{ConstrainedLinReg}} and a data.frame.
#' @param object Model of class \code{\link{ConstrainedLinReg}}
#' @param newdata data.frame containing all variables that appear in the model
#' formula
#' @return Numeric vector of predictions. For observations with missing
#' values on the explanatory variables, a prediction of \code{NA} is returned.
#' @export
setMethod("predict",
          "ConstrainedLinReg",
          function(object, newdata) {
            newdataPrep <- prepDatForPredict(object@formula, newdata)
            
            X <- model.matrix(object = object@formula, data = newdataPrep)
            beta <- colMeans(getBetaMatrix(object, object@hasIntercept))
            
            pred <- as.vector(X %*% beta)
            pred <- napredict(attributes(newdataPrep)$na.action, pred)
            
            return(pred)
          }
)


#' Exclude rows with missing data on predictor variables
#' @description Rows with missing values on predictor variables are excluded.
#' An unused column for the dependent variable is added to avoid errors.
#' @details A column of ones for the dependent variable is added. Otherwise
#' \code{\link[stats]{model.matrix}} tries to take it from the formula's
#' environment, which is the original data. This usually results in an error due
#' to unequal variable length. This column is however not used.
#' @param formula Model formula
#' @param newdata data.frame containing all variables that appear in the model
#' @return Object of class \code{\link[stats]{na.exclude}}
prepDatForPredict <- function(formula, newdata) {
  predictors <- all.vars(formula, functions = FALSE)[-1]
  datRes <- na.exclude(newdata[, predictors, drop = FALSE])
  
  # Add column of ones for dependent variable.
  datRes[[all.vars(formula, functions = FALSE)[1]]] <- rep(1, nrow(datRes))
  
  return(datRes)
}