Skip to contents

Get Effective AIC Range for Gibbs Temperature Scaling

Source: R/get_effective_aic_range.R
get_effective_aic_range.Rd

Calculates the expected delta AIC range for a given top percentile using a continuous relationship based on log odds ratio.

Usage

get_effective_aic_range(top_percentile)

Arguments

top_percentile

Numeric percentage (0-100) of top simulations

Value

Numeric effective range for delta AIC normalization

Details

The effective range represents the expected delta AIC span for the top X% of models under typical (non-dispersed) conditions.

Uses a heuristic that maps percentiles to Burnham & Anderson's AIC thresholds. Based on their interpretation:

  • 1% → \(\Delta\) AIC < 2 (substantial support, ~37% relative likelihood)

  • 5% → \(\Delta\) AIC about 4 (boundary of substantial/less support)

  • 10% → \(\Delta\) AIC about 7 (upper end of "considerably less support")

  • 20% → \(\Delta\) AIC about 10 (boundary to "essentially no support")

This heuristic is used only by plot_effective_range. It does not drive calibration weighting: run_MOSAIC(), grid_search_best_subset and optimize_ensemble_subset saturate \(\Delta\) AIC at a fixed 4 for the best subset regardless of the subset's percentile.

Examples

get_effective_aic_range(1)   # 2.0
#> [1] 2
get_effective_aic_range(5)   # 4.0
#> [1] 4
get_effective_aic_range(10)  # 7.0
#> [1] 7