Train Cox Proportional Hazards model for supervised PCA

Main and utility functions for training the Cox PH model.

coxTrain_fun(x, y, censoring.status, s0.perc = NULL)

Arguments

x

A "tall" pathway data frame (\(p \times n\)).
y

A response vector of follow-up / event times.
censoring.status

A censoring vector.
s0.perc

A stabilization parameter. This is an optional argument to each of the functions called internally. Defaults to NULL.

Value

A list containing:

  • tt : The scaled p-dimensional score vector: each value has been divided by the respective standard deviation plus the fudge value.

  • numer : The original p-dimensional score vector. From the internal .coxscor function.

  • sd : The standard deviations of the scores. From the internal .coxvar function.

  • fudge : A regularization scalar added to the standard deviation. If s0.perc is supplied, fudge = quantile(sd, s0.perc).

Details

See https://web.stanford.edu/~hastie/Papers/spca_JASA.pdf, Section 5, for a description of Supervised PCA applied to survival data. The internal utility functions defined in this file (.coxscor, .coxvar, and .coxstuff) are not called anywhere else, other than in the coxTrain_fun function itself. Therefore, we do not document these functions.

NOTE: No missing values allowed.

Examples

  # DO NOT CALL THIS FUNCTION DIRECTLY.
  # Use SuperPCA_pVals() instead

if (FALSE) {
  p <- 500
  n <- 50

  x_mat <- matrix(rnorm(n * p), nrow = p, ncol = n)
  x_df <- data.frame(x_mat)
  time_int <- rpois(n, lambda = 365 * 2)
  obs_logi <- sample(
    c(FALSE, TRUE),
    size = n,
    replace = TRUE,
    prob = c(0.2, 0.8)
  )

  coxTrain_fun(
    x = x_df,
    y = time_int,
    censoring.status = !obs_logi
  )
}