Area under the curve of binomial GAM predictions vs. observations

Fits a Quasibinomial GAM model y ~ s(x) (y ~ x if x is non-numeric) with the binomial response y (values 0 and 1) and the numeric, character or factor predictor x using mgcv::gam() and returns the area under the ROC curve between the observed responses and the model predictions (see score_auc()).

Cases are weighted with case_weights() to prevent issues arising from class imbalance.

Supports cross-validation via the arguments arguments cv_training_fraction (numeric between 0 and 1) and cv_iterations (integer between 1 and n) introduced via ellipsis (...). See preference_order() for further details.

Usage

f_binomial_gam(df, ...)

Arguments

df

(required, dataframe) with columns:

• "x": (numeric, character, factor) predictor.

• "y" (integer) binomial response with unique values 0 and 1.

...

(optional) Accepts the arguments cv_training_fraction (numeric between 0 and 1) and cv_iterations (integer between 1 and Inf) for cross validation.

Value

numeric or numeric vector: AUC

See also

Other preference_order_functions: f_binomial_glm(), f_binomial_rf(), f_categorical_rf(), f_count_gam(), f_count_glm(), f_count_rf(), f_numeric_gam(), f_numeric_glm(), f_numeric_rf(), preference_order()

Examples

data(vi_smol, package = "spatialData")

df <- data.frame(
  y = vi_smol[["vi_binomial"]],
  x = vi_smol[["swi_max"]]
)

#no cross-validation
f_binomial_gam(df = df)
#> [1] 0.7947985

#cross-validation
f_binomial_gam(
  df = df,
  cv_training_fraction = 0.5,
  cv_iterations = 10
  )
#>  [1] 0.7471018 0.7926659 0.7864337 0.7561480 0.7648046 0.7835627 0.7821146
#>  [8] 0.7813673 0.7369469 0.7756643

#categorical predictor
df <- data.frame(
  y = vi_smol[["vi_binomial"]],
  x = vi_smol[["koppen_zone"]]
)

f_binomial_gam(df = df)
#> [1] 0.9331935