Skip to contents

Projection uncertainty settings

Source: R/uncertainty.R
uncertainty_settings.Rd

A projected map is a surface of point estimates, and a point estimate on its own invites more confidence than it has earned. Two different things can be wrong with a cell, and they need separate answers because a cell can be badly affected by one and untouched by the other:

Usage

uncertainty_settings(config)

Arguments

config

a config list, as returned by load_config()

Value

NULL when off, otherwise a list with method, replicates, level, and novelty

Details

  • The fit could have been different. Refit on slightly different data and the surface moves. ensemble_spread() measures how much.

  • The cell may be somewhere the model has never seen. novelty_surface() is the one that catches this, and the spread cannot: a narrow interval means the members agree, not that they are right. Where the covariates are off the end of the training data, members can agree perfectly and all be extrapolating.

The two do not track each other, and neither substitutes for the other. They often move together — a resampled member's training range differs from the full model's, so cells near the edge tend to be both novel and unstable — but a confident prediction over a novel cell is exactly the case a spread-only map reports as trustworthy. Read both.

Off by default. On a real grid over a decade the extra prediction passes are a real multiplier, and most runs are iterations that only want the mean surface.

projection:
  uncertainty: true          # or the block below, for the non-defaults
  uncertainty:
    method: folds            # or: bootstrap
    replicates: 100          # bootstrap only
    level: 0.9               # interval width
    novelty: true

See also

novelty_surface(), ensemble_spread()

Examples

config <- load_config(
  system.file("configs/mock_test.yaml", package = "taupatch")
)
uncertainty_settings(config)                      # NULL: off by default
#> NULL

config$projection$uncertainty <- TRUE
uncertainty_settings(config)
#> $method
#> [1] "folds"
#> 
#> $replicates
#> [1] 100
#> 
#> $level
#> [1] 0.9
#> 
#> $novelty
#> [1] TRUE
#> 

config$projection$uncertainty <- list(method = "bootstrap", level = 0.95)
uncertainty_settings(config)
#> $method
#> [1] "bootstrap"
#> 
#> $replicates
#> [1] 100
#> 
#> $level
#> [1] 0.95
#> 
#> $novelty
#> [1] TRUE
#>