Draw an interactive multiverse of analyses — ggmultiverse • ggextreme

Draws a specification curve: every reasonable analysis of the same question, one per row of data, sorted by its estimate, with its confidence interval, above a grid that marks the choices each analysis made. The primary analysis stays marked, the reference value is drawn, and beside each choice sits the median estimate of the analyses that made it, so the choices that move the result stand out.

Usage

ggmultiverse(
  data,
  estimate,
  lower,
  upper,
  decisions,
  primary = NULL,
  ratio = NULL,
  ylab = "Estimate",
  title = NULL,
  caption = NULL,
  family = "Lato"
)

Arguments

data: A data frame with one row per analysis.
estimate, lower, upper: Bare columns of data with each estimate and its confidence interval.
decisions: Names of the columns of data that hold the choices, such as the outcome definition, the adjustment set and the model.
primary: Optional logical expression of the columns of data, or a row number, marking the primary analysis.
ratio: Whether the estimates are ratios, drawn on a log scale around 1. Defaults to TRUE when every lower limit is positive.
ylab: Label of the estimate axis, such as "Odds ratio".
title, caption: Title above the plot and note below it.
family: Font family. The package ships Lato and registers it on load.

Value

An object of class ggmultiverse, which prints as an interactive widget. Use graph_widget(), graph_plot() or graph_save() for the widget, a static ggplot or a file. The field influence holds the median estimate for every choice.

Details

In the widget, hovering over an analysis shows its estimate and every choice behind it. Dragging across the curve selects a run of analyses and says which choices they share. Clicking a choice in the grid keeps only the analyses that made it, and several choices can be combined; the primary analysis is never hidden.

The share of analyses with intervals that exclude the reference value is a description of this set of analyses, not a probability that the effect is real, and the analyses are not independent.

Examples

specs <- expand.grid(outcome = c("Primary", "Broad"),
                     adjustment = c("Minimal", "Standard", "Extended"),
                     model = c("Logistic", "Log-binomial"),
                     stringsAsFactors = FALSE)
set.seed(1)
log_or <- -0.3 + 0.12 * (specs$outcome == "Broad") +
  0.1 * match(specs$adjustment, c("Minimal", "Standard", "Extended")) +
  rnorm(nrow(specs), 0, 0.03)
specs$or <- exp(log_or)
specs$lo <- exp(log_or - 1.96 * 0.09)
specs$hi <- exp(log_or + 1.96 * 0.09)
ggmultiverse(specs, or, lo, hi, decisions = c("outcome", "adjustment", "model"),
             primary = outcome == "Primary" & adjustment == "Standard" & model == "Logistic",
             ylab = "Odds ratio")