R/untransform_delim.R

In plmmr: Penalized Linear Mixed Models for Correlated Data

#' Untransform coefficient values back to the original scale for **file-backed** data
#'
#' This function unwinds the initial standardization of the data to obtain
#' coefficient values on their original scale. It is called by plmm_format().
#'
#' @param std_scale_beta The estimated coefficients on the standardized scale
#' @param p  The number of columns in the original design matrix
#' @param std_X_details A list with 3 elements describing the standardized design matrix BEFORE rotation; this should have elements 'scale', 'center', and 'ns'
#' @param use_names Logical: should names be added? Defaults to TRUE. Set to FALSE inside of `cvf()` helper, as 'ns' will vary within CV folds.
#' @keywords internal
#'
#' @returns a matrix of estimated coeffcients, 'beta_vals', that is on the scale of the original data.
untransform_delim <- function(std_scale_beta, p, std_X_details, plink_flag,
                              use_names = TRUE) {

  # goal: reverse the PRE-ROTATION standardization #
  # partition the values from Step 1 into intercept and non-intercept parts
  a <- std_scale_beta[1, , drop = FALSE] # this is the intercept
  b <- std_scale_beta[-1, , drop=FALSE]

  # initialize beta with zeros; nrow = # of predictors, ncol = # of lambda values
  # this will create columns of zeros for betas corresponding to singular columns
  untransformed_beta <- Matrix::Matrix(0,
                                       nrow = (p + 1), # + 1 is for the intercept; see note below
                                       ncol = ncol(std_scale_beta), # this is the number of lambda values
                                       sparse = TRUE)


  # next, unscale the beta values for non-singular, non-intercept columns
  # NB: this requires the details of standardization (centering/scaling values
  # and indices of nonsingular columns). The details of how these details are
  # passed around varies depending on whether data are stored filebacked,
  # hence, the division into cases below:
  if (length(std_X_details$ns) == length(std_X_details$scale)){
    # case 1: ns and center/scale values have same length
    untransformed_b2 <- sweep(x = b,
                              MARGIN = 1,
                              STATS = std_X_details$scale,
                              FUN = "/")

    # fill in the un-transformed values
    untransformed_beta[std_X_details$ns+1,] <- untransformed_b2 # again, the + 1 is for the intercept
    cp <- apply(X = untransformed_b2, 2, function(c){crossprod(std_X_details$center, c)})
    untransformed_beta[1,] <- a - cp
  } else {
    # case 2: ns and center/scale values **do not** have same length
    # (this will often be the case in cross-validation, where features can
    # become constant in a given fold)
    untransformed_b2 <- sweep(x = b[std_X_details$ns,],
                              MARGIN = 1,
                              STATS = std_X_details$scale[std_X_details$ns],
                              FUN = "/")

    # fill in the un-transformed values
    untransformed_beta[std_X_details$ns+1,] <- untransformed_b2 # again, the + 1 is for the intercept
    cp <- apply(X = untransformed_b2, 2, function(c){crossprod(std_X_details$center[std_X_details$ns], c)})
    untransformed_beta[1,] <- a - cp
  }

  if (use_names) {
    rownames(untransformed_beta) <- c("(Intercept)",
                                      std_X_details$X_colnames)
  }


  # Final step: return un-transformed beta values
  return(untransformed_beta)
}