Jackknife Variable Importance Analysis — maxent

Performs a leave-one-variable-out jackknife analysis following the procedure in Java Maxent's Runner.jackknifeGain(). For each environmental variable, two additional models are trained:

Leave-one-out: the variable is excluded and the model is trained on all remaining variables.
Only-one: the model is trained using only that single variable.

The resulting training gain and AUC are compared to the full model to assess variable importance.

Usage

maxent_jackknife(
  env_vals,
  sample_indices,
  num_points,
  types = c("linear", "quadratic", "hinge"),
  n_hinges = 15L,
  max_iter = 500L,
  categorical = NULL,
  bias_weights = NULL
)

Arguments

env_vals: Named list of numeric vectors (one per environmental variable, length = total points: background + occurrences).
sample_indices: Integer vector: 0-based indices of occurrence samples within env_vals.
num_points: Integer: total number of points (background + occurrences).
types: Character vector of feature types (default c("linear", "quadratic", "hinge")).
n_hinges: Integer: number of hinge knots (default 15).
max_iter: Integer: maximum training iterations (default 500).
categorical: Character vector of variable names that are categorical (default NULL).
bias_weights: Optional numeric vector of per-point bias weights (default NULL).

Value

A data.frame with columns:

variable: Variable name.
gain_without: Training gain of the model with this variable excluded.
gain_only: Training gain of the model using only this variable.
gain_full: Training gain of the full model (same for all rows).

Examples

if (FALSE) { # \dontrun{
env <- list(
  temp  = rnorm(110),
  precip = rnorm(110),
  elev  = rnorm(110)
)
jk <- maxent_jackknife(env, sample_indices = 100:109,
                        num_points = 110L)
print(jk)
} # }