Skip to contents

Bootstrap for RESIT

Source: R/lingam_resit_bootstrap.r
lingam_resit_bootstrap.Rd

Bootstrap for RESIT

Usage

lingam_resit_bootstrap(
  X,
  n_sampling,
  regressor = "gam",
  alpha = 0.01,
  prior_knowledge = NULL,
  seed = NULL,
  verbose = TRUE,
  parallel = FALSE,
  n_cores = NULL
)

Arguments

X

Numeric matrix (n_samples x n_features)

n_sampling

Number of bootstrap iterations

regressor

Nonlinear regressor, passed to lingam_resit()

alpha

Significance level of the HSIC pruning test, passed to lingam_resit()

prior_knowledge

Prior-knowledge matrix, passed to lingam_resit()

seed

Random seed (NULL allowed)

verbose

Whether to display progress (logical)

parallel

Whether to use parallel processing (logical)

n_cores

Number of cores to use (integer, NULL allowed)

Value

A BootstrapResult (list); see lingam_direct_bootstrap() for the query helpers that operate on it (get_probabilities(), get_causal_direction_counts(), get_directed_acyclic_graph_counts(), get_causal_order_stability()).

Details

total_effects is always NULL: RESIT is a nonlinear method for which total causal effects are undefined, so there is no compute_total_effects argument and get_total_causal_effects() raises its usual "no total effects" error. (The Python implementation instead stores an all-zero total-effects array; storing nothing is deliberate, so the zeros cannot be mistaken for estimated effects.)

causal_orders is populated (RESIT estimates a full causal order), so get_causal_order_stability() works with the returned object.

Each iteration runs O(ncol(X)^2) nonlinear regressions plus HSIC tests, so a bootstrap is substantially slower than the linear variants; keep n_sampling modest. When parallel = TRUE, a user-supplied regressor function is serialized to the PSOCK workers; any package it uses must be referenced with the pkg::fun form inside the function body.

Examples

# \donttest{
if (requireNamespace("mgcv", quietly = TRUE)) {
  nonlinear <- generate_resit_sample(n = 300, seed = 1)

  bs <- lingam_resit_bootstrap(nonlinear$data,
    n_sampling = 3L,
    seed = 42
  )
  get_probabilities(bs)
}
#> Bootstrap: 3 iterations, RESIT (sequential)
#>   iteration 1 / 3
#> Completed in 1.1 seconds.
#>      [,1] [,2] [,3] [,4]
#> [1,]    0    0    0    0
#> [2,]    1    0    0    0
#> [3,]    1    1    0    0
#> [4,]    0    0    1    0
# }