r.gauss.pardag: Generate a Gaussian Causal Model Randomly

Description

Generate a random Gaussian causal model. Parameters specifying the connectivity as well as coefficients and error terms of the corresponding linear structural equation model can be specified. The observational expectation value of the generated model is always 0, meaning that no interception terms are drawn.

Usage

r.gauss.pardag(p, prob, top.sort = FALSE, normalize = FALSE,
               lbe = 0.1, ube = 1, neg.coef = TRUE, labels = as.character(1:p),
               lbv = 0.5, ubv = 1)

Arguments

the number of nodes.

prob

probability of connecting a node to another node.

top.sort

logical indicating whether the output graph should be topologically sorted, meaning that arrows always point from lower to higher node indices.

normalize

logical indicating whether weights and error variances should be normalized such that the diagonal of the corresponding observational covariance matrix is 1.

lbe, ube

lower and upper bounds of the absolute values of edge weights.

neg.coef

logical indicating whether negative edge weights are also admissible.

labels

(optional) character vector of variable (or “node”) names.

lbv, ubv

lower and upper bound on error variances of the noise terms in the structural equations.

Value

An object of class "'>GaussParDAG".

Details

The underlying directed acyclic graph (DAG) is generated by drawing an undirected graph from an Erd<U+0151>s-R<U+00E9>nyi model orienting the edges according to a random topological ordering drawn uniformly from the set of permutations of p variables. This means that any two nodes are connected with (the same) probability prob, and that the connectivity of different pairs of nodes is independent.

A Gaussian causal model can be represented as a set of linear structural equations. The regression coefficients of the model can be represented as "edge weights" of the DAG. Edge weights are drawn uniformly and independently from the interval between lbe and ube; if neg.coef = TRUE, their sign is flipped with probability 0.5. Error variances are drawn uniformly and independently from the interval between lbv and ubv.

If normalize = TRUE, the edge weights and error variances are normalized in the end to ensure that the diagonal elements of the observational covariance matrix are all 1; the procedure used is described in Hauser and B<U+00FC>hlmann (2012). Note that in this case the error variances and edge weights are no longer guaranteed to lie in the specified intervals after normalization.

References

P. Erd<U+0151>s and A. R<U+00E9>nyi (1960). On the evolution of random graphs. Publications of the Mathematical Institute of the Hungarian Academy of Sciences 5, 17--61.

A. Hauser and P. B<U+00FC>hlmann (2012). Characterization and greedy learning of interventional Markov equivalence classes of directed acyclic graphs. Journal of Machine Learning Research 13, 2409--2464.

Examples

Run this code

# NOT RUN {
set.seed(307)

## Plot some random DAGs
if (require(Rgraphviz)) {
  ## Topologically sorted random DAG
  myDAG <- r.gauss.pardag(p = 10, prob = 0.2, top.sort = TRUE)
  plot(myDAG)

  ## Unsorted DAG
  myDAG <- r.gauss.pardag(p = 10, prob = 0.2, top.sort = FALSE)
  plot(myDAG)
}

## Without normalization, edge weigths and error variances lie within the
## specified borders
set.seed(307)
myDAG <- r.gauss.pardag(p = 10, prob = 0.4,
  lbe = 0.1, ube = 1, lbv = 0.5, ubv = 1.5, neg.coef = FALSE)
B <- myDAG$weight.mat()
V <- myDAG$err.var()
any((B > 0 & B < 0.1) | B > 1)
any(V < 0.5 | V > 1.5)

## After normalization, edge weights and error variances are not necessarily
## within the specified range, but the diagonal of the observational covariance
## matrix consists of ones only
set.seed(308)
myDAG <- r.gauss.pardag(p = 10, prob = 0.4, normalize = TRUE,
  lbe = 0.1, ube = 1, lbv = 0.5, ubv = 1.5, neg.coef = FALSE)
B <- myDAG$weight.mat()
V <- myDAG$err.var()
any((B > 0 & B < 0.1) | B > 1)
any(V < 0.5 | V > 1.5)
diag(myDAG$cov.mat())
# }

Run the code above in your browser using DataCamp Workspace