R/GLMnet.R

In riskRegression: Risk Regression Models and Prediction Scores for Survival Analysis with Competing Risks

#' @title Fitting GLMnet for use with predictRisk
#'
#' @description Fit GLMnet models via a formula and a data set for use with \code{\link{predictRisk}}.
#' @name GLMnet
#'
#' @param formula A formula.
#' @param data The data on which to fit the model. 
#' @param lambda The tuning parameters for GLMnet. If set to NULL, then it the parameters are chosen for you.
#' @param cv Whether to use cross-validation or not. Default is TRUE.
#' @param alpha The elasticnet mixing parameter. See the ?glmnet for more details.
#' @param nfolds Number of folds for cross-validation. Default is 10.
#' @param type.measure loss to use for cross-validation. Default is deviance.
#' @param family passed to \code{glmnet}. Defaults for binary outcome to \code{"binomial"} and for survival to \code{"cox"}.
#' @param \dots Additional arguments that are passed on to the glmnet.
#' @export
GLMnet <- function(formula,
                   data,
                   lambda=NULL,
                   cv=TRUE,
                   alpha = 1,
                   nfolds = 10,
                   type.measure = "deviance",
                   family,
                   ...){
  requireNamespace(c("glmnet","prodlim"))
  tt <- all.vars(update(formula,".~1"))
  if (length(tt) != 1){
    if (missing(family)) family <- "cox"
    strata.num = start = status = NULL
    EHF = prodlim::EventHistory.frame(formula,data,unspecialsDesign = TRUE,specials = NULL)
    stopifnot(attr(EHF$event.history,"model")[[1]] == "survival")
    # blank Cox object needed for predictions
    data = data.frame(cbind(EHF$event.history,EHF$design))
    bl_cph <- coxph(Surv(time,status)~1,data=data,x=1,y=1)
    bl_obj <- coxModelFrame(bl_cph)[]
    bl_obj[,strata.num:=0]
    bl_obj[,order := 1:.N]
    data.table::setorder(bl_obj, strata.num,stop,start,-status)
    sorted_x_train = EHF$design[bl_obj[,order],] ## save this for calculating Breslow estimator when doing predictions
    ## bl_obj$linear_predictor = c(exp(predict(fit,newx = EHF$design,type = "link", s=lambda)))
    design <- EHF$design[bl_obj$order,]
    if (!cv){
      fit <- glmnet::glmnet(x=EHF$design,
                            y=EHF$event.history,
                            lambda=lambda,
                            alpha= alpha,
                            family=family,
                            ...)
    }
    else {
      fit <- glmnet::cv.glmnet(x=EHF$design,
                               y=EHF$event.history,
                               lambda=lambda,
                               nfolds=nfolds,
                               type.measure = type.measure,
                               alpha=alpha,
                               family=family,
                               ...)
      lambda <- fit$lambda
    }
  }
  else {
    if (missing(family)) family <- "binomial"
    sorted_x_train=bl_obj=terms=design = NULL
    y  <- data[[tt[1]]]
    x <- model.matrix(formula, data=data)
    if (!cv){
      fit <- glmnet::glmnet(x=x,
                            y=y,
                            lambda=lambda,
                            alpha=alpha,
                            family=family,
                            ...)
    }
    else {
      fit <- glmnet::cv.glmnet(x=x,
                               y=y,
                               lambda=lambda,
                               nfolds=nfolds,
                               type.measure =type.measure,
                               alpha=alpha,
                               family=family,
                               ...)
      lambda <- fit$lambda
    }
  }
  out = list(fit = fit,
             surv_info = bl_obj,
             design = design,
             call = match.call(),
             terms = terms(formula),
             sorted_x_train=sorted_x_train,
             cv=cv,
             lambda = lambda)
  class(out) = "GLMnet"
  out
}