causalDisco (version 1.0.1)

tpc_run: Run the TPC Algorithm for Causal Discovery

Description

Run a tier-aware variant of the PC algorithm that respects background knowledge about a partial temporal order. Supply the temporal order via a knowledge object.

Usage

tpc_run(
  data = NULL,
  knowledge = NULL,
  alpha = 0.05,
  test = reg_test,
  suff_stat = NULL,
  method = "stable.fast",
  na_method = "none",
  orientation_method = "conservative",
  directed_as_undirected = FALSE,
  varnames = NULL,
  num_cores = 1,
  ...
)

Value

A Disco object (a list with a caugi and a knowledge object).

Arguments

data

A data frame with the observed variables.

knowledge

A knowledge object created with knowledge(), encoding tier assignments and optional forbidden/required edges. This is the preferred way to provide temporal background knowledge.

alpha

The alpha level used as the per-test significance threshold for conditional independence testing.

test

A conditional independence test. The default reg_test uses a regression-based information-loss test. Another available option is cor_test which tests for vanishing partial correlations. User-supplied functions may also be used; see details for the required interface.

suff_stat

A sufficient statistic. If supplied, it is passed directly to the test and no statistics are computed from data. Its structure depends on the chosen test.

method

Skeleton construction method, one of "stable", "original", or "stable.fast" (default). See pcalg::skeleton() for details.

na_method

Handling of missing values, one of "none" (default; error on any NA), "cc" (complete-case analysis), or "twd" (test-wise deletion).

orientation_method

Conflict-handling method when orienting edges. Currently only the conservative method is available.

directed_as_undirected

Logical; if TRUE, treat any directed edges in knowledge as undirected during skeleton learning. This is due to the fact that pcalg does not allow directed edges in fixedEdges or fixedGaps. Default is FALSE.

varnames

Character vector of variable names. Only needed when data is not supplied and all information is passed via suff_stat.

num_cores

Integer number of CPU cores to use for parallel skeleton learning.

...

Additional arguments passed to pcalg::skeleton() during skeleton construction.

Details

Any independence test implemented in pcalg may be used; see pcalg::pc(). When na_method = "twd", test-wise deletion is performed: for cor_test, each pairwise correlation uses complete cases; for reg_test, each conditional test performs its own deletion. If you supply a user-defined test, you must also provide suff_stat.

Temporal or tiered knowledge enters in two places:

• during skeleton estimation, candidate conditioning sets are pruned so they do not contain variables that are strictly after both endpoints;

• during orientation, any cross-tier edge is restricted to point forward in time.

Examples

# Load data
data(tpc_example)

# Build knowledge
kn <- knowledge(
  tpc_example,
  tier(
    child ~ tidyselect::starts_with("child"),
    youth ~ tidyselect::starts_with("youth"),
    old ~ tidyselect::starts_with("old")
  )
)

# Recommended route using disco
my_tpc <- tpc(engine = "causalDisco", test = "fisher_z", alpha = 0.05)

disco(tpc_example, my_tpc, knowledge = kn)

# or using my_tpc directly

my_tpc <- my_tpc |> set_knowledge(kn)
my_tpc(tpc_example)

# Using tpc_run() directly

tpc_run(tpc_example, knowledge = kn, alpha = 0.01)