R/grad.R

In benkeser/survtmle: Compute Targeted Minimum Loss-Based Estimates in Right-Censored Survival Settings

#' Gradient for Logistic Regression
#'
#' @description A function that computes the gradient of the for a logistic
#'  regression model. Used by \code{\link[stats]{optim}} on occasion.
#'
#' @param beta A vector of coefficients in a logistic glm
#' @param X The design matrix
#' @param Y The outcome
#'
#' @importFrom stats plogis
#'
#' @return Numeric vector of the gradient of the parameter vector.
grad <- function(beta, Y, X) {
  pi <- stats::plogis(X %*% beta)
  pi[pi == 0] <- .Machine$double.neg.eps
  pi[pi == 1] <- 1 - .Machine$double.neg.eps
  gr <- crossprod(X, Y - pi)
  return(-gr)
}

#' Gradient for Logistic Regression with Offsets
#'
#' @description A function that computes the gradient of the for a logistic
#'  regression model with an offset term. Used by \code{\link[stats]{optim}} on
#'  occasion.
#'
#' @param beta A vector of coefficients in a logistic glm
#' @param H The covariate matrix
#' @param offset The offset vector
#' @param Y The outcome
#'
#' @importFrom stats plogis
#'
#' @return Numeric vector of the gradient of the parameter vector.
grad_offset <- function(beta, Y, H, offset = NULL) {
  pi <- stats::plogis(cbind(offset, H) %*% c(1, beta))
  pi[pi == 0] <- .Machine$double.neg.eps
  pi[pi == 1] <- 1 - .Machine$double.neg.eps
  gr <- crossprod(H, Y - pi)
  return(-gr)
}