Interactive network plots • ggextreme

A network meta-analysis compares several treatments at once by combining trials that each compare a few of them. The network plot shows which comparisons the evidence covers: every treatment is a node, and every pair of treatments compared directly in at least one trial is joined by a line. Judging whether the trials are similar enough to combine means going back to the arm level data, the baseline characteristics and outcomes of each arm. ggnma() puts that data behind the plot. Hovering over a node or a line shows its arms side by side, in the manner of a trial’s baseline table, and clicking it opens the full table with every column.

library(ggextreme)

A first network

The package includes arm level data from five randomized trials in moderate to severe plaque psoriasis, compiled by Phillippo et al. for their multilevel network meta-regression.

net <- ggnma(
  psoriasis_nma,
  study = study,
  treatment = treatment,
  n = n,
  group = class,
  legend_title = "Class",
  title = "Treatments for plaque psoriasis"
)
net

Hover over the line between placebo and secukinumab 150 mg to compare the four trials that make that comparison, arm by arm, and click it for every column. Hover over or click a node to see every arm on that treatment. The references at the bottom of each panel are linked.

The data

ggnma() takes long data with one row per study arm, the shape netmeta::pairwise() and multinma::set_agd_arm() start from:

head(psoriasis_nma[c("study", "treatment", "n", "pasi75_r", "age")])
#>     study          treatment   n pasi75_r  age
#> 1   CLEAR        Ustekinumab 339      265 44.6
#> 2   CLEAR Secukinumab 300 mg 337      304 45.2
#> 3 ERASURE            Placebo 248       11 45.4
#> 4 ERASURE Secukinumab 150 mg 245      174 44.9
#> 5 ERASURE Secukinumab 300 mg 245      200 44.9
#> 6 FEATURE            Placebo  59        0 46.5

Three columns place each arm: study, treatment and, optionally, group for the drug class. Every other column becomes a row of the panel tables, in the order given, so choosing what to show is a matter of selecting columns first. n is optional; when given, node area follows the total number of participants on each treatment, and the hover cards count participants as well as studies.

Rows are named from each column’s label attribute when it has one, as set by the ‘labelled’, ‘Hmisc’ or ‘haven’ packages, and otherwise from the column name:

attr(psoriasis_nma$age, "label")
#> [1] "Mean age (years)"

Text that is the same for every arm of a study, such as the reference here, is listed once per study under the table rather than repeated in every column. Text that contains a URL or a DOI is linked.

The hover card holds a compact version of the table. By default it shows the first four columns; hover picks them by name, so the card can carry the sample size, the outcome and the baseline characteristic that matters most for the question at hand:

ggnma(
  psoriasis_nma, study, treatment, n = n, group = class,
  hover = c("n", "pasi75_r", "pasi_w0", "prior_systemic"),
  legend = FALSE
)

hover = character(0) gives a card that only lists the studies. The click panel always shows every column.

What the plot encodes

Nodes are treatments, placed on a circle starting at the top and running clockwise in the order of the factor levels of treatment, or in order of first appearance. positions places them by hand.
Node area follows the number of participants on each treatment when n is given.
Line width follows the number of studies that make the comparison.
A multi-arm study adds a line for every pair of its treatments, and a shaded polygon joining them. The counts are available on the object:

net$edges
#>                 from                 to studies    n
#> 1            Placebo         Etanercept       1  652
#> 2            Placebo Secukinumab 150 mg       4 1386
#> 3            Placebo Secukinumab 300 mg       4 1385
#> 4         Etanercept Secukinumab 150 mg       1  653
#> 5         Etanercept Secukinumab 300 mg       1  653
#> 6        Ustekinumab Secukinumab 300 mg       1  676
#> 7 Secukinumab 150 mg Secukinumab 300 mg       4 1383

Multi-arm studies

Lines alone cannot show that three treatments were compared within one trial rather than in three separate ones. A shaded polygon joins the treatments of every study with more than two arms, and studies that compare the same set share one polygon. Here ERASURE, FEATURE and JUNCTURE each compare placebo with both doses of secukinumab, and FIXTURE adds etanercept:

net$multiarm
#>                                                    treatments arms
#> 1             Placebo, Secukinumab 150 mg, Secukinumab 300 mg    3
#> 2 Placebo, Etanercept, Secukinumab 150 mg, Secukinumab 300 mg    4
#>                      studies
#> 1 ERASURE, FEATURE, JUNCTURE
#> 2                    FIXTURE

Each polygon has its own hover card and panel, with the arms of every study it stands for. multiarm = FALSE leaves the shading out.

Placing nodes by hand

positions takes one row per treatment with x and y coordinates, y pointing up. Any units will do: the layout is scaled to fit the plot, keeping its shape, and labels point away from its middle. This one puts placebo at the top, the active comparators at the sides and the two secukinumab doses along the bottom:

positions <- data.frame(
  treatment = c("Placebo", "Etanercept", "Ustekinumab",
                "Secukinumab 150 mg", "Secukinumab 300 mg"),
  x = c(1, 0, 2, 0.5, 1.5),
  y = c(1, 0, 0, -1, -1)
)
ggnma(psoriasis_nma, study, treatment, n = n, group = class,
      positions = positions, legend = FALSE)

Colors

Without group, every node takes the first color of race_palette(), or the single color in palette. With group, nodes are colored by class and a legend is drawn. A named palette sets colors by class, and classes left out keep their default:

ggnma(
  psoriasis_nma, study, treatment, n = n, group = class,
  palette = c(Placebo = "#9A9A9A"),
  legend_title = "Class"
)

Where the evidence for a comparison comes from

Give a netmeta fit on the same network, or the result of netmeta::netcontrib() for it, as contributions, and the widget gains a menu of every comparison. Picking one widens and colors each line by the share of that network estimate flowing through it, labels the shares and lists them under the plot, while the lines it does not use fade.

pw <- meta::pairwise(treat = treatment, event = pasi75_r, n = pasi75_n,
                     studlab = study, data = psoriasis_nma, sm = "OR")
fit <- netmeta::netmeta(pw, common = FALSE)
ggnma(psoriasis_nma, study, treatment, n = n, group = class,
      legend_title = "Class", contributions = fit)

No trial compares ustekinumab with placebo, so its estimate flows entirely through the other comparisons, most of it through the one trial of ustekinumab against secukinumab.

Other forms

As with causal diagrams, the plot prints as a widget in the RStudio viewer, in R Markdown and Quarto documents, and on ‘pkgdown’ sites. graph_widget() returns the htmlwidget for ‘shiny’, graph_plot() the underlying ggplot for a manuscript, and graph_save() a single .html file or a static .png:

graph_save(net, "network.html")
graph_save(net, "network.png", res = 300)

Source of the example data

The arm level data are those compiled by Phillippo (2019) and distributed with the ‘multinma’ package, from the published reports of CLEAR, ERASURE, FEATURE, FIXTURE and JUNCTURE; see ?psoriasis_nma. They were analyzed in Phillippo et al. (2020).