Skip to contents

Spatial cross-validation across mesh cutoff values

Source: R/mesh-cv.R
sdmTMB_mesh_cv.Rd

Compares alternative mesh resolutions for an sdmTMB model using spatial block cross-validation. Spatial folds are created with blockCV::cv_spatial() and each mesh is evaluated with sdmTMB::sdmTMB_cv() using the same folds, so that predictive performance (sum of hold-out log-likelihoods and AIC) can be compared on an equal footing.

This is a parameterised version of an exploratory workflow: all model settings are inherited from the fitted object, and nothing about the data is assumed beyond the coordinate and response columns.

Usage

sdmTMB_mesh_cv(
  object,
  mesh_cutoffs,
  k = 3,
  block_size = NULL,
  response = NULL,
  xy_cols = c("X", "Y"),
  coord_multiplier = 1000,
  object_crs = NULL,
  control = NULL,
  n_workers = 1,
  future_seed = TRUE,
  silent = TRUE
)

Arguments

object

A fitted sdmTMB::sdmTMB() model. Model settings (formula, family, spatial and spatiotemporal structure, time column) are inherited.

mesh_cutoffs

Numeric vector of mesh cutoff values to compare.

k

Integer. Number of cross-validation folds. Default 3.

block_size

Numeric. Spatial block size in CRS units (e.g. metres). If NULL (default) a heuristic based on the data extent is used.

response

Character. Name of the response column used to build presence/absence blocks for blockCV::cv_spatial(). If NULL (default), taken from the model formula's left-hand side.

xy_cols

Character vector of the coordinate columns. Default c("X", "Y").

coord_multiplier

Numeric factor to convert xy_cols to CRS units. Default 1000.

object_crs

Coordinate reference system. If NULL, taken from the model mesh.

control

An sdmTMB::sdmTMBcontrol() object for the CV fits. If NULL, inherited from object.

n_workers

Integer. Number of parallel workers. Default 1.

future_seed

Logical or integer. Future-compatible random seed.

silent

Logical. Suppress fitting output? Default TRUE.

Value

A data frame (tibble) of class sdmTMB_mesh_cv with one row per cutoff: cutoff, crashed, converged, sanity, sum_loglik (the summed hold-out log-likelihood from sdmTMB::sdmTMB_cv(); higher is better), mean_aic, and the derived deltaNLL, deltaNLL_pr, deltaAIC, deltaAIC_pr. deltaNLL/deltaAIC are relative to the best-fitting cutoff (0 = best; lower is better, mirroring AIC). The list of sdmTMB::sdmTMB_cv() objects is attached as the "cv" attribute.

See also

sdmTMB::sdmTMB_cv(), blockCV::cv_spatial(), compare_meshes()

Author

Mikko Vihtakari

Examples

# \donttest{
# Requires the 'blockCV' package.
mesh <- sdmTMB::make_mesh(sdmTMB::pcod, c("X", "Y"), cutoff = 20)
fit <- sdmTMB::sdmTMB(
  data = sdmTMB::pcod, formula = density ~ 0 + as.factor(year),
  time = "year", mesh = mesh, family = sdmTMB::tweedie(link = "log")
)
cv <- sdmTMB_mesh_cv(fit, mesh_cutoffs = c(15, 30), k = 3)
#> Running fits with `future.apply()`.
#> Set a parallel `future::plan()` to use parallel processing.
#> Running fits with `future.apply()`.
#> Set a parallel `future::plan()` to use parallel processing.
#> ℹ `ln_tau_O` is an internal parameter affecting `sigma_O`
#> ℹ `sigma_O` is the spatial standard deviation
#> ℹ `ln_tau_E` is an internal parameter affecting `sigma_E`
#> ℹ `sigma_E` is the spatiotemporal standard deviation
#> ℹ `ln_kappa` is an internal parameter affecting `range`
#> ℹ `range` is the distance at which data are effectively independent
#> ℹ `ln_tau_O` is an internal parameter affecting `sigma_O`
#> ℹ `sigma_O` is the spatial standard deviation
#> ℹ `ln_tau_E` is an internal parameter affecting `sigma_E`
#> ℹ `sigma_E` is the spatiotemporal standard deviation
#> ℹ `ln_kappa` is an internal parameter affecting `range`
#> ℹ `range` is the distance at which data are effectively independent
#> ℹ `ln_tau_O` is an internal parameter affecting `sigma_O`
#> ℹ `sigma_O` is the spatial standard deviation
#> ℹ `ln_tau_E` is an internal parameter affecting `sigma_E`
#> ℹ `sigma_E` is the spatiotemporal standard deviation
#> ℹ `ln_kappa` is an internal parameter affecting `range`
#> ℹ `range` is the distance at which data are effectively independent
cv
#> # A tibble: 2 × 10
#>   cutoff crashed converged sanity sum_loglik mean_aic deltaNLL deltaNLL_pr
#>    <int> <lgl>   <lgl>     <lgl>       <dbl>    <dbl>    <dbl>       <dbl>
#> 1     30 FALSE   TRUE      FALSE      -7613.    8781.       0          0  
#> 2     15 FALSE   TRUE      TRUE       -9819.    8644.    2206.        29.0
#> # ℹ 2 more variables: deltaAIC <dbl>, deltaAIC_pr <dbl>
plot(cv)

# }