hib_cv_model: Cross-validate a fitted hibiiscus mixed-effects model

This function re-fits models on different subsets of the data, and estimates a range of error metrics allowing for evaluation of model predictive performance.

Usage

hib_cv_model(
  mod_obj,
  db_obj,
  fold_type,
  custom_col = NULL,
  random_seed = NULL,
  re_form = list(NA),
  backtransform = NULL,
  use_original_mesh = NULL
)

Arguments

mod_obj: One of; HibModelLME4, HibModelINLA or HibModelBRMS.
db_obj: A HibDatabase object, or, if the above model used multiple data sources, a list of HibDatabase objects.
fold_type: A character indicating the type of cross-validation; one of: "Random-2", "Random-3", "Random-5", "Random-10", "Source_ID", "SS", "Country", "UN_subregion", "UN_region", "Biome", "Realm", "Custom".
custom_col: A character indicating the column in the model data (sites data.frame) to use to define cross-validation folds, when fold_type == "Custom".
random_seed: A numeric indicating the random seed used to define random cross-validation folds - only used if 'Random-X' in fold_type.
re_form: A list indicating the random effects variables to be included in the cross-validation/insample predicitons - if multiple items supplied, each option will be used in a separate cross-validated prediction. Values can be either NA (no random effects), NULL (all random effects), or a rhs formula specifying particular random effects to include.
Note: only test set data points with random-effect values present in the training data will have predictions if re_form is not NA.
backtransform: A function to convert the model prediction back to the original scale of interest. The inverse of any transformation used in model fitting.
use_original_mesh: A logical value indicating whether cross-validation of spatial INLA models should use the original mesh for all CV folds. Computationally this is slightly more efficient, but will impact CV performance metrics as 'withheld' sites will still exist in the mesh.

Value

HibCV

Details

This function allows for model evaluation using multiple cross-validation strategies, and multiple random effects structures. It returns a HibCV object, which includes the original model, insample and CV performance for multiple evaluation metrics, model coefficients estimated per cross-validation fold.

Examples

if (FALSE) { # \dontrun{
  cv_out <- hib_cv_model(mod_obj, db_obj, fold_type = "Random-10")
} # }