Plot Method for Multi-Source KL-Integrated Cox Elastic-Net Models

Produces a line plot of model performance (loss or C-index) as a function of the transfer-learning shrinkage parameter \(\eta\) for each external source in a coxkl_enet.multi object. Each source is displayed as a separate line with a distinct color and linetype.

Usage

# S3 method for class 'coxkl_enet.multi'
plot(
  x,
  test_z = NULL,
  test_time = NULL,
  test_delta = NULL,
  test_stratum = NULL,
  criteria = c("loss", "CIndex"),
  ...
)

Arguments

x

An object of class "coxkl_enet.multi", as returned by coxkl_enet.multi.

test_z

Optional numeric matrix of test predictors of dimension n_test x p. If NULL (default), training data stored inside each source fit are used for evaluation.

test_time

Optional numeric vector of test survival times of length n_test. Must be provided together with test_z and test_delta when evaluating on external test data.

test_delta

Optional numeric vector of test event indicators of length n_test. Must be provided together with test_z and test_time when evaluating on external test data.

test_stratum

Optional vector of stratum indicators of length n_test for stratified Cox models. Ignored if NULL (default).

criteria

Character string specifying the performance metric to plot. Either "loss" (default, negative log partial likelihood scaled by sample size) or "CIndex" (Harrell's concordance index).

...

Currently unused. Reserved for future extensions.

Value

A ggplot object. The plot can be further customized with standard ggplot2 layers and themes.

Details

For each valid source fit stored in x$source_fits, the function extracts the p x n_eta coefficient matrix integrated_stat.betahat_best, where each column corresponds to one value of \(\eta\). It then calls test_eval() on every column to compute the chosen performance metric, and overlays the resulting curves on a single ggplot2 figure.

If all four test arguments (test_z, test_time, test_delta, test_stratum) are NULL, evaluation is performed on the training data embedded in each source fit object. This is useful for a quick in-sample diagnostic but may give optimistic estimates of performance.

Colors are assigned automatically via hue_pal and linetypes cycle through "solid", "dashed", "dotdash", "longdash", and "twodash" to remain distinguishable when printed in grayscale.

See also

coxkl_enet.multi for fitting the multi-source model, plot.coxkl for the analogous plot method for single-source "coxkl" objects.

Examples

if (FALSE) { # \dontrun{
fit <- coxkl_enet.multi(
  z         = z_train,
  delta     = delta_train,
  time      = time_train,
  beta_list = list(beta_ext1, beta_ext2, beta_ext3),
  etas      = seq(0, 1, by = 0.1)
)

# In-sample diagnostic (uses training data stored in each source fit)
plot(fit, criteria = "CIndex")

# Out-of-sample evaluation on a held-out test set
plot(fit,
     test_z     = z_test,
     test_time  = time_test,
     test_delta = delta_test,
     criteria   = "loss")
} # }