Skip to contents

Prediction Stability Map

Source: R/borg_stability_map.R
borg_stability_map.Rd

Creates a continuous spatial map of prediction variance across cross-validation folds. Unlike borg_stability() which returns summary statistics, this function produces per-location stability scores showing where predictions are consistent versus volatile across different train/test partitions.

Usage

borg_stability_map(
  model,
  data,
  new = NULL,
  target,
  coords,
  formula = NULL,
  fit_fun = NULL,
  folds = NULL,
  v = 10,
  seed = 42
)

Arguments

model

A fitted model with a predict() method, or a model-fitting function.

data

Data frame with predictors, target, and coordinates.

new

Data frame of prediction locations. If NULL, uses data.

target

Character. Target variable name.

coords

Character vector of length 2. Coordinate column names.

formula

Model formula. Required if model is a function.

fit_fun

Function. Model fitting function. If NULL, attempts to refit from model's call.

folds

A borg_cv or borg_disc_cv object, or a fold list. If NULL, generates spatial CV folds automatically.

v

Integer. Number of folds if generating automatically. Default: 10.

seed

Integer. Random seed. Default: 42.

Value

A data frame with class "borg_stability_map" containing:

pred_mean

Mean prediction across folds

pred_sd

Standard deviation of predictions across folds

pred_cv

Coefficient of variation (|sd/mean|)

pred_range

Range of predictions (max - min)

n_folds_predicted

Number of folds where this point was in the test set (or could be predicted)

x, y

Coordinates

Has print() and autoplot() methods.

Details

How it works

For each CV fold:

  1. Refit the model on the training set.

  2. Predict on all locations (not just test set).

  3. Store per-location predictions.

Then compute per-location statistics across folds. Locations with high pred_sd are unstable — the prediction changes substantially depending on which data is included in training.

Interpretation

High instability at a location can indicate:

  • The location is in a data-sparse region (near AOA boundary)

  • The model is overfitting to nearby training points

  • The underlying relationship changes spatially (non-stationarity)

Examples

set.seed(42)
d <- data.frame(x = runif(100, 0, 50), y = runif(100, 0, 50))
d$z <- sin(d$x / 5) + rnorm(100, sd = 0.5)
model <- lm(z ~ x + y, data = d)
sm <- borg_stability_map(model, d, target = "z", coords = c("x", "y"),
                           v = 5)
sm