Visualizing networks with cograph

Drawing a fitted network

A fitted psychnets network is a weighted graph of nodes and signed edges. The picture of that graph is where its structure becomes readable: which nodes connect, how strongly, and with what sign. Every estimator in psychnets returns an object of class c("psychnet", "cograph_network"), so a fitted network moves into the cograph drawing engine with no conversion. The same model is then rendered with different layouts, node aesthetics, edge styling, thresholds, themes, and export settings, all from one object.

This vignette shows the plotting workflow from the default graph to a figure ready for a paper. cograph::splot() is the drawing verb throughout. The chunks that call it are guarded, so they run only when cograph is installed.

The data and the fit

The bundled SRL_GPT data hold 300 learners scored on five self-regulated -learning constructs (CSU, IV, SE, SR, TA). ebic_glasso() estimates a Gaussian graphical model whose edges are partial correlations selected by the extended Bayesian information criterion.

fit <- ebic_glasso(SRL_GPT)
fit
#> <psychnet> glasso network
#>   nodes: 5   edges: 10   (undirected)
#>   lambda: 0.00861   gamma: 0.5
#>   optimality (KKT residual): 2.21e-10

The default graph

plot() on a psychnet object delegates to cograph::splot(), so a bare plot(fit) draws the network. Each node is a construct and each edge is a partial correlation between two constructs given the other three. The colour of an edge encodes the sign of that partial correlation and the width encodes its magnitude, so a wide edge of one colour is a strong positive association and a wide edge of the other colour is a strong negative one.

plot(fit)

Calling cograph::splot() directly draws the same graph and opens its full argument surface for the customization below.

cograph::splot(fit)

Layouts

The layout argument sets the algorithm that places the nodes. A circular layout fixes the nodes on a ring, which keeps positions comparable between figures; a spring layout places connected nodes near one another, which makes clusters visible. The seed argument fixes the random start of a stochastic layout so the figure is reproducible.

op <- par(mfrow = c(1, 2), mar = c(1, 1, 3, 1))
cograph::splot(fit, layout = "circle", title = "circle")
cograph::splot(fit, layout = "spring", seed = 11, title = "spring")

par(op)

Nodes and labels

The node arguments set fill, border, label size, and shape. The scale_nodes_by argument sizes each node by a centrality measure and node_size_range sets the smallest and largest radius, so node area reads as structural importance. Sizing the nodes by strength makes the most connected construct the largest on the page.

cograph::splot(
  fit,
  layout = "circle",
  scale_nodes_by = "strength",
  node_size_range = c(4, 12),
  node_fill = c("#4C78A8", "#F58518", "#54A24B", "#B279A2", "#E45756"),
  node_border_color = "white",
  node_border_width = 2,
  label_size = 0.9,
  title = "Node size scaled by strength"
)

Edges

For a signed psychometric network, the edge encoding carries the substantive result. The edge_positive_color and edge_negative_color arguments set the two sign colours and edge_width_range sets the mapping from magnitude to width. The threshold argument hides edges below an absolute weight, which thins a dense graph down to its strongest associations, and edge_labels prints the weight on each retained edge.

cograph::splot(
  fit,
  layout = "circle",
  threshold = 0.05,
  edge_width_range = c(0.5, 5),
  edge_positive_color = "#2A9D8F",
  edge_negative_color = "#E76F51",
  edge_labels = TRUE,
  edge_label_size = 0.7,
  edge_label_bg = "white",
  title = "Thresholded edge weights"
)

cograph::plot_edge_weights() draws the distribution of the edge weights on its own, which reads the spread of associations and the location of the strong edges without the graph layout.

cograph::plot_edge_weights(fit)

Bootstrap diagnostics

The bootstrap diagnostics are computed in psychnets and drawn by their own base-R plot methods, covered fully in the companion vignette. net_boot() resamples the data and refits the network on each resample, and plot() on its result defaults to the edge-weight confidence intervals.

set.seed(1)
bs <- net_boot(SRL_GPT, method = "glasso", n_boot = 250, cores = 1)
plot(bs)

The same bootstrap object holds the retained draws for the pairwise edge-difference test. Under type = "edge_diff", plot() draws the significance-box matrix of which edges differ from one another.

plot(bs, type = "edge_diff")

difference_test() returns the pairwise differences as a tidy table, and plot() under style = "forest" draws each difference as a point with its confidence interval when the effect sizes matter more than the box display.

plot(difference_test(bs, type = "edge"), style = "forest")

Themes

The theme argument applies a coordinated set of colour and styling defaults without changing the fitted model, so the same network is redrawn in a house style or a colourblind-safe palette from one call.

op <- par(mfrow = c(1, 2), mar = c(1, 1, 3, 1))
cograph::splot(fit, theme = "minimal", title = "minimal")
cograph::splot(fit, theme = "colorblind", title = "colorblind")

par(op)

Group networks

Passing group = to psychnet() fits one network per level of a grouping column and returns the collection. cograph::splot() draws the collection as a grid that shares one layout, so a node sits in the same position in every panel and the panels are read side by side.

group_fit <- psychnet(grouped_srl, group = "source", method = "glasso")
cograph::splot(group_fit, layout = "circle", psych_styling = TRUE)

Export

The filename, width, height, and res arguments write the figure to a file at a chosen size and resolution when it is ready to save.

cograph::splot(fit, layout = "circle", filename = "network.png",
               width = 8, height = 8, res = 300)

The splot arguments at a glance