Interactive forest plots • ggextreme

A forest plot summarizes a meta-analysis in one figure, but the judgments behind it, what each trial measured, how it was run and how far it can be trusted, live in an extraction sheet the reader never sees. ggmeta() draws the forest plot of a fitted model and keeps that sheet with it. Hovering over a study shows its effect, weight and chosen columns; clicking it opens its full record; and risk of bias judgments sit beside each study as traffic lights.

library(ggextreme)

A first forest plot

The example is the review by Chiarito et al. (2020) of P2Y12 inhibitors against aspirin for secondary prevention in atherosclerosis, which ships with metafor. It includes their risk of bias assessment with the RoB 2 tool.

dat <- metafor::escalc(
  measure = "OR",
  ai = p2y12.mi, n1i = p2y12.total,
  ci = aspirin.mi, n2i = aspirin.total,
  data = metadat::dat.chiarito2020,
  slab = paste(study, year)
)
dat <- dat[!is.na(dat$yi), ]
dat$p2y12 <- paste0(dat$p2y12.mi, "/", dat$p2y12.total)
dat$aspirin <- paste0(dat$aspirin.mi, "/", dat$aspirin.total)

fit <- metafor::rma(yi, vi, data = dat)

ggmeta(
  fit,
  columns = c("P2Y12 inhibitor" = "p2y12", Aspirin = "aspirin"),
  rob = c(R = "rob.R", D = "rob.D", Mi = "rob.Mi", Me = "rob.Me",
          S = "rob.S", Overall = "rob.overall"),
  favors = c("Favors P2Y12 inhibitor", "Favors aspirin"),
  title = "Myocardial infarction"
)

Hover over a study for its effect, weight and risk of bias in each domain, and click it for every column of the data. Hover over the diamond for the heterogeneity statistics and the prediction interval.

The input

ggmeta() takes a fitted model rather than raw numbers, so the estimates are the ones the analysis reports:

an rma.uni fit from metafor::rma(), without moderators, or
a meta object from the meta package, such as meta::metabin() or meta::metagen().

The extraction record comes from the data stored in the fit, which metafor keeps when the model is fitted with a data argument. Pass data to use a separate sheet instead, with one row per study in the order of the model.

argument	effect
`columns`	columns printed beside the study labels, named to set the headers
`rob`	risk of bias columns, one per domain, drawn as traffic lights
`hover`	columns shown in each study’s hover card; defaults to `columns`
`favors`	labels for the two sides of the null line
`xlim`	axis limits; intervals that run past them end in an arrow
`exponentiate`	show ratio measures on the ratio scale, the default for them

Risk of bias

rob names the columns that hold judgments, one per domain, in the order they should appear. Judgments are matched by their wording, so the RoB 2, RoB 1 and ROBINS-I scales all work as written:

wording	light
low, low risk	green, +
some concerns, moderate	amber, −
unclear	amber, ?
high, serious	red, ×
critical	dark red, !
no information	gray, ?

Each light carries its own hover card, and a key of the judgments in use is drawn above the plot. A wording outside these is refused with a message naming it, rather than drawn in a color that might mislead.

Cumulative meta-analysis

With cumulative = TRUE, each row shows the pooled estimate from the studies up to and including it. Sort the data by year before fitting the model to see how the evidence built up over time. The BCG vaccine trials, published between 1948 and 1980, show the pooled estimate moving as each trial was added.

bcg <- metafor::escalc(
  measure = "RR", ai = tpos, bi = tneg, ci = cpos, di = cneg,
  data = metadat::dat.bcg, slab = paste(author, year)
)
bcg <- bcg[order(bcg$year), ]
bcg_fit <- metafor::rma(yi, vi, data = bcg)

ggmeta(bcg_fit, cumulative = TRUE, favors = c("Favors BCG", "Favors control"))

animate_meta() replays the same sequence for a talk or a supplement. Each trial fades in as it is added and the pooled diamond eases to its new position, the same motion as the bar chart race, with the year shown behind the plot:

animate_meta(ggmeta(bcg_fit), "bcg.gif", time = bcg$year)

Other forms

The forest 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, and graph_save() a single .html file or a static .png at the plot’s natural size.