pcalg (version 2.4-3)

pdsep: Estimate Final Skeleton in the FCI algorithm

Description

Estimate the final skeleton in the FCI algorithm (Spirtes et al, 2000), as described in Steps 2 and 3 of Algorithm 3.1 in Colombo et al. (2012). The input of this function consists of an initial skeleton that was estimated by the PC algorithm (Step 1 of Algorithm 3.1 in Colombo et al. (2012)).

Given the initial skeleton, all unshielded triples are considered and oriented as colliders when appropriate. Then, for all nodes x in the resulting partially directed graph G, Possible-D-SEP(x,G) is computed, using the function qreach. Finally, for any edge y-z that is present in G, conditional independence between Y and Z is tested given all subsets of Possible-D-SEP(y,G) and all subsets of Possible-D-SEP(z,G). These tests are done at level alpha, using indepTest. If the pair of nodes is judged to be independent given some set S, then S is recorded in sepset(y,z) and sepset(z,y) and the edge y-z is deleted. Otherwise, the edge remains and there is no change to sepset.

Usage

pdsep(skel, suffStat, indepTest, p, sepset, alpha, pMax, m.max = Inf, pdsep.max = Inf, NAdelete = TRUE, unfVect = NULL, biCC = FALSE, verbose = FALSE)

Arguments

skel

Graph object returned by skeleton.

suffStat

Sufficient statistic: A list containing all necessary elements for making conditional independence decisions using function indepTest.

indepTest

Predefined function for testing conditional independence. The function is internally called as indepTest(x,y,S,suffStat) for testing conditional independence of x and y given S. Here, x and y are node numbers of the adjacency matrix, S is a (possibly empty) vector of node numbers of the adjacency matrix and suffStat is a list containing all relevant elements for making conditional independence decisions. The return value of indepTest is the p-value of the test for conditional independence.

p

Number of variables.

sepset

List of length p; each element of the list contains another list of length p. The element sepset[[x]][[y]] contains the separation set that made the edge between x and y drop out. This object is thought to be obtained from a pcAlgo-object or fciAlgo-object.

alpha

Significance level for the individual conditional independence tests.

pMax

Matrix with the maximal p-values of conditional independence tests in a previous call of skeleton, pc or fci which produced G. This object is thought to be obtained from a pcAlgo-object or fciAlgo-object.

m.max

Maximum size of the conditioning sets that are considered in the conditional independence tests.

pdsep.max

Maximum size of Possible-D-SEP for which subsets are considered as conditioning sets in the conditional independence tests. If the nodes x and y are adjacent in the graph and the size of Possible-D-SEP(x,G)\ x,y, is bigger than pdsep.max, the edge is simply left in the graph. Note that if pdsep.max is less than Inf, the final PAG is typically a supergraph of the one computed with pdsep.max = Inf, because fewer tests may have been performed in the former.

NAdelete

If indepTest returns NA and this option is TRUE, the corresponding edge is deleted. If this option is FALSE, the edge is not deleted.

unfVect

Vector containing numbers that encode the unfaithful triple (as returned by pc.cons.intern). This is needed in the conservative FCI.

biCC

Logical; if TRUE, only nodes on paths between nodes a and c are considered to be in sepset(a,c). This uses biconnected components, see biConnComp from RBGL.

verbose

Logical indicating that detailed output is to be provided.

Value

Details

To make the code more efficient, we only perform tests that were not performed in the estimation of the initial skeleton.

Note that the Possible-D-SEP sets are computed once in the beginning. They are not updated after edge deletions, in order to make sure that the output of the algorithm does not depend on the ordering of the variables (see also Colombo and Maathuis (2014)).

References

P. Spirtes, C. Glymour and R. Scheines (2000). Causation, Prediction, and Search, 2nd edition. The MIT Press.

D. Colombo, M.H. Maathuis, M. Kalisch and T.S. Richardson (2012). Learning high-dimensional directed acyclic graphs with latent and selection variables. Annals of Statistics 40, 294--321.

D. Colombo and M.H. Maathuis (2014).Order-independent constraint-based causal structure learning. Journal of Machine Learning Research 15 3741-3782.

See Also

qreach to find Possible-D-SEP(x,G); fci.

Examples

p <- 10
## generate and draw random DAG:
set.seed(44)
myDAG <- randomDAG(p, prob = 0.2)

## generate 10000 samples of DAG using gaussian distribution
library(RBGL)
n <- 10000
d.mat <- rmvDAG(n, myDAG, errDist = "normal")

## estimate skeleton
indepTest <- gaussCItest
suffStat <- list(C = cor(d.mat), n = n)
alpha <- 0.01
skel <- skeleton(suffStat, indepTest, alpha=alpha, p=p)

## prepare input for pdsep
sepset <- skel@sepset
pMax <- skel@pMax

## call pdsep to find Possible-D-Sep and enhance the skeleton
pdsepRes <- pdsep(skel@graph, suffStat, indepTest, p, sepset, alpha,
                  pMax, verbose = TRUE)
## call pdsep with biconnected components to find Possible-D-Sep and enhance the skeleton
pdsepResBicc <- pdsep(skel@graph, suffStat, indepTest, p, sepset, alpha,
                      pMax, biCC= TRUE, verbose = TRUE)