Relative-importance network (LMG / Shapley)

Builds a directed network in which the edge predictor -> outcome is the predictor's LMG (Shapley) share of the outcome node's regression R-squared. Equivalent in purpose to relaimpo::calc.relimp(type = "lmg") applied nodewise / bootnet's "relimp" default, pure base R and self-certified via lmg_certificate().

Usage

relimp_network(
  data = NULL,
  cor_matrix = NULL,
  cor_method = c("pearson", "spearman", "kendall", "auto"),
  max_nodes = 21L,
  normalized = FALSE,
  na_method = c("pairwise", "listwise"),
  labels = NULL
)

Arguments

data

Numeric data frame or matrix (rows = observations). Optional if cor_matrix is supplied.

cor_matrix

Optional correlation matrix.

cor_method

Correlation when data is supplied: "pearson" (default), "spearman", "kendall", or "auto" (polychoric/polyserial; see cor_auto()).

max_nodes

Refuse to run above this many nodes (the cost grows as 2^(p-1) per node). Default 21.

normalized

If TRUE, rescale each outcome's incoming importance shares to sum to 1, matching bootnet/relaimpo's normalized reporting. The default FALSE keeps raw LMG shares that sum to the outcome R-squared.

na_method

Missing-data handling when data is supplied: "pairwise" (default) or "listwise". See ebic_glasso().

labels

Optional node labels.

Value

A psychnet object whose $weights is the directed importance matrix (weights[k, j] = importance of k for outcome j), with $r2 (per-node full-model R-squared), $cor_matrix, and $kkt (the decomposition residual). With normalized = FALSE (default) each outcome's incoming shares sum to its R-squared (colSums($weights) == $r2); with normalized = TRUE they are rescaled to sum to 1, $raw_importance holds the unscaled shares, and $kkt (like lmg_certificate()) is computed from those raw shares.

Examples

S <- 0.4^abs(outer(1:5, 1:5, "-"))
relimp_network(cor_matrix = S)
#> <psychnet> relimp network
#>   nodes: 5   edges: 20   (directed)
#>   optimality (KKT residual): 5.55e-17