Supervised Linear Encoding of Factors via the Cox Component of a Joint Model

step_lencode_joint_model creates a specification of a recipe step that encodes categorical predictors into a numeric score: the log hazard ratio of each level relative to the reference level (the first level, encoded as 0), taken from the Cox component of a joint_model fit to that factor alone. Only the Cox component's .pred_risk_score is used, so the encoding is a penalized Cox log hazard ratio, like that of step_lencode_coxnet; it is not a composite that blends the survival, status and follow-up time components of the joint model, and engine does not affect it. Levels that bake() has not seen, or missing values, are encoded as 0. If a model can't be fit, a warning is raised and the factor's levels are all encoded as 0.

Usage

step_lencode_joint_model(
  recipe,
  ...,
  outcome = c("time", "status"),
  role = "predictor",
  trained = FALSE,
  engine = "glmnet",
  penalty = 0.05,
  mapping = NULL,
  skip = FALSE,
  id = recipes::rand_id("lencode_joint_model")
)

# S3 method for class 'step_lencode_joint_model'
required_pkgs(x, ...)

Arguments

recipe

A recipe object.

...

One or more selector functions to choose variables to be encoded.

outcome

Character vector naming the survival outcome columns in the training data, as for step_lencode_coxnet.

role

Role for the encoded variables. Default is "predictor".

trained

A logical indicating whether the step has been trained.

engine

Has no effect on the encoding and is ignored when the step is prepped: only the Cox component of the joint model is used, and it does not depend on the engine. Any value other than the default "glmnet" gives a warning. Kept so that existing code keeps running.

penalty

Penalty of the Cox component. Default is 0.05.

mapping

A named list of level-to-score mappings generated during prep().

skip

Logical; skip step when baking? Default is FALSE.

id

A unique identifier for the step.

x

A step_lencode_joint_model object.

Value

An updated version of recipe.

Missing values

A missing value in a factor counts as the reference level, exactly as in step_lencode_coxnet: in prep() it is recoded to the first level before the model is fit, so its row stays in the fit, and in bake() it is encoded as 0.

The training data are encoded with their own fit

prep() fits the encoding on the training rows, and bake(new_data = NULL), or baking the training data, encodes those same rows with the coefficients fitted to them: the encoding is not out-of-fold (cross-fitted). The encoded training column therefore carries the outcome, so it looks more predictive in the training data than it is for new subjects, and a model fit to it over-trusts it. The effect grows with the number of levels relative to the number of events, so it is worst for factors with many or rare levels, whose coefficients are fitted to a handful of subjects each: a factor with 100 levels and no relation to the outcome can show a clearly better than chance concordance in the baked training data and chance level on new data. To limit it, use a large penalty (most levels then shrink to 0) or pool rare levels before encoding (for example with recipes::step_other()). Do not evaluate a model on data that it was encoded with: estimate performance on subjects the step did not see in prep(), for example by resampling the whole recipe, so that every prep() sees only the analysis set, and not by encoding once and splitting afterwards.