Skip to contents

GRaSP Algorithm for Causal Discovery

Source: R/grasp.R
grasp.Rd

Run the Greedy Relaxations of the Sparsest Permutation algorithm for causal discovery using one of several engines.

Usage

grasp(engine = "tetrad", score, test, alpha = 0.05, ...)

Arguments

engine

Character; which engine to use. Must be one of:

"tetrad"

Tetrad Java library.

score

Character; name of the scoring function to use.

test

Character; name of the conditional‐independence test.

alpha

Numeric; significance level for the CI tests.

...

Additional arguments passed to the chosen engine (e.g. score and algorithm parameters).

Details

For specific details on the supported scores, and parameters for each engine, see:

Recommendation

While it is possible to call the function returned directly with a data frame, we recommend using disco(). This provides a consistent interface and handles knowledge integration.

Value

A function that takes a single argument data (a data frame). When called, this function returns a list containing:

  • knowledge A Knowledge object with the background knowledge used in the causal discovery algorithm. See knowledge() for how to construct it.

  • caugi A caugi::caugi object (of class PDAG) representing the learned causal graph from the causal discovery algorithm.

References

Lam, W.-Y., Andrews, B., & Ramsey, J. (2022). Greedy Relaxations of the Sparsest Permutation Algorithm. In The 38th Conference on Uncertainty in Artificial Intelligence.

See also

Other causal discovery algorithms: boss(), boss_fci(), fci(), ges(), gfci(), grasp_fci(), gs(), iamb-family, pc(), rfci(), sp_fci(), tfci(), tges(), tpc()

Examples

data(tpc_example)

# Requires Tetrad to be installed
if (verify_tetrad()$installed && verify_tetrad()$java_ok) {
  # Recommended path using disco()
  grasp_tetrad <- grasp(
    engine = "tetrad",
    test = "fisher_z",
    score = "sem_bic",
    alpha = 0.05
  )
  disco(tpc_example, grasp_tetrad)

  # or using grasp_tetrad directly
  grasp_tetrad(tpc_example)
}
#> <Disco UNKNOWN: 6 nodes | 6 edges>
#>   nodes: child_x2, child_x1, youth_x4, youth_x3, oldage_x6, oldage_x5
#>   edges: child_x2---child_x1, child_x2-->oldage_x5, child_x2---youth_x4
#>          oldage_x5-->oldage_x6, youth_x3-->oldage_x5, youth_x4-->oldage_x6

#### With tier knowledge ####
if (verify_tetrad()$installed && verify_tetrad()$java_ok) {
  kn <- knowledge(
    tpc_example,
    tier(
      child ~ tidyselect::starts_with("child"),
      youth ~ tidyselect::starts_with("youth"),
      oldage ~ tidyselect::starts_with("oldage")
    )
  )

  # Recommended path using disco()
  grasp_tetrad <- grasp(
    engine = "tetrad",
    test = "fisher_z",
    score = "sem_bic",
    alpha = 0.05
  )
  disco(tpc_example, grasp_tetrad, knowledge = kn)

  # or using grasp_tetrad directly
  grasp_tetrad <- grasp_tetrad |> set_knowledge(kn)
  grasp_tetrad(tpc_example)
}
#> <Disco UNKNOWN: 6 nodes | 6 edges>
#>   nodes: child_x2, child_x1, youth_x4, youth_x3, oldage_x6, oldage_x5
#>   edges: child_x2---child_x1, child_x2-->oldage_x5, child_x2-->youth_x4
#>          oldage_x5-->oldage_x6, youth_x3-->oldage_x5, youth_x4-->oldage_x6

# With all algorithm arguments specified
if (verify_tetrad()$installed && verify_tetrad()$java_ok) {
  grasp_tetrad <- grasp_fci(
    engine = "tetrad",
    test = "poisson_prior",
    score = "rank_bic",
    alpha = 0.05,
    depth = 3,
    stable_fas = FALSE,
    max_disc_path_length = 5,
    covered_depth = 3,
    singular_depth = 2,
    nonsingular_depth = 2,
    ordered_alg = TRUE,
    raskutti_uhler = TRUE,
    use_data_order = FALSE,
    num_starts = 3
  )
  disco(tpc_example, grasp_tetrad)
}
#> <Disco PAG: 6 nodes | 6 edges>
#>   nodes: child_x2, child_x1, youth_x4, youth_x3, oldage_x6, oldage_x5
#>   edges: child_x2o-ochild_x1, child_x2o->oldage_x5, child_x2o-oyouth_x4
#>          oldage_x5-->oldage_x6, youth_x3o->oldage_x5, youth_x4-->oldage_x6