diff --git a/NAMESPACE b/NAMESPACE
index 3405d737..0429f1b5 100644
--- a/NAMESPACE
+++ b/NAMESPACE
@@ -145,6 +145,7 @@ export(pareto_k_ids)
 export(pareto_k_influence_values)
 export(pareto_k_table)
 export(pareto_k_values)
+export(plot_loo_difference)
 export(pointwise)
 export(print_dims)
 export(pseudobma_weights)
diff --git a/R/loo_difference_plot.R b/R/loo_difference_plot.R
new file mode 100644
index 00000000..dc60f3f5
--- /dev/null
+++ b/R/loo_difference_plot.R
@@ -0,0 +1,213 @@
+#' Compare models across domains
+#'
+#' The LOO difference plot shows how the ELPD of two different models
+#' changes when a predictor is varied. This can be useful for identifying
+#' opportunities for model stacking or expansion. Pointwise differences
+#' are computed as `loo_1 - loo_2`, so positive values indicate better
+#' predictive performance for `loo_1`.
+#'
+#' @param y A vector of observations.
+#' @param loo_1,loo_2 Objects returned by [loo()].
+#' @param group An optional grouping variable with the same length as `y`.
+#'   Points are colored according to group membership.
+#' @param size,alpha Point size and opacity passed to [ggplot2::geom_point()].
+#' @param jitter Amount of horizontal jitter passed to
+#'   [ggplot2::position_jitter()].
+#' @param sort_by_group If `TRUE`, observations are ordered by `group`
+#'   and the x-axis is replaced by a sequential index. The supplied `y` values
+#'   are therefore not used as x coordinates. Plotting by index can be useful
+#'   when categories have very different sample sizes. To control the group
+#'   order, supply `group` as a factor with levels in the desired order.
+#' @param label_threshold Optional nonnegative threshold for labeling
+#'   observations. Observations for which the absolute pointwise ELPD
+#'   difference exceeds this value are labeled. If `NULL`, no observations
+#'   are labeled.
+#' @param labels Optional vector of labels with the same length as `y`, used
+#'   for observations selected by `label_threshold`. If `NULL`, observation
+#'   indices are used.
+#'
+#' @template bayesvis-reference
+#'
+#' @return A [ggplot2::ggplot()] object.
+#'
+#' @examples
+#' # Artificial example
+#' log_lik <- example_loglik_matrix()
+#' shift <- seq(-0.5, 0.5, length.out = ncol(log_lik))
+#' log_lik_2 <- sweep(log_lik, 2, shift, FUN = "+")
+#'
+#' loo_1 <- loo(log_lik)
+#' loo_2 <- loo(log_lik_2)
+#'
+#' plot_loo_difference(
+#'   seq_len(ncol(log_lik)),
+#'   loo_1,
+#'   loo_2
+#' )
+#'
+#' # Label observations with large pointwise ELPD differences
+#' plot_loo_difference(
+#'   seq_len(ncol(log_lik)),
+#'   loo_1,
+#'   loo_2,
+#'   label_threshold = 0.3
+#' )
+#'
+#' # Create interspersed groups, then sort them in the plot
+#' group <- rep(c("A", "A", "A", "B"), length.out = ncol(log_lik))
+#'
+#' plot_loo_difference(
+#'   seq_len(ncol(log_lik)),
+#'   loo_1,
+#'   loo_2,
+#'   group = group,
+#'   sort_by_group = TRUE
+#' )
+#'
+#' @export
+plot_loo_difference <-
+  function(
+    y,
+    loo_1,
+    loo_2,
+    group = NULL,
+    size = 1,
+    alpha = 1,
+    jitter = 0,
+    sort_by_group = FALSE,
+    label_threshold = NULL,
+    labels = NULL
+  ) {
+    if (!requireNamespace("ggplot2", quietly = TRUE)) {
+      stop(
+        "Please install 'ggplot2' to use `plot_loo_difference()`.",
+        call. = FALSE
+      )
+    }
+
+    checkmate::assert_flag(sort_by_group)
+    loo_compare_checks(nlist(loo_1, loo_2))
+
+    # elpd_diffs(a, b) computes b - a
+    elpd_diff <- elpd_diffs(loo_2, loo_1)
+
+    checkmate::assert_atomic_vector(
+      y,
+      len = length(elpd_diff)
+    )
+
+    if (!is.null(group)) {
+      checkmate::assert_atomic_vector(
+        group,
+        len = length(y),
+        any.missing = FALSE
+      )
+    }
+
+    if (!is.null(labels)) {
+      checkmate::assert_atomic_vector(
+        labels,
+        len = length(y)
+      )
+
+      if (is.null(label_threshold)) {
+        stop(
+          "`label_threshold` must be supplied when `labels` is supplied.",
+          call. = FALSE
+        )
+      }
+    }
+
+    if (!is.null(label_threshold)) {
+      checkmate::assert_number(
+        label_threshold,
+        lower = 0,
+        finite = TRUE
+      )
+    }
+
+    if (!is.null(label_threshold) && is.null(labels)) {
+      labels <- seq_along(y)
+    }
+
+    if (sort_by_group) {
+      if (is.null(group)) {
+        stop(
+          "`group` must be supplied when `sort_by_group = TRUE`.",
+          call. = FALSE
+        )
+      }
+
+      ordering <- order(group)
+      elpd_diff <- elpd_diff[ordering]
+      group <- group[ordering]
+
+      if (!is.null(labels)) {
+        labels <- labels[ordering]
+      }
+
+      y <- seq_along(elpd_diff)
+    }
+
+    plot_data <- data.frame(
+      y = y,
+      elpd_diff = elpd_diff
+    )
+
+    if (!is.null(group)) {
+      plot_data$group <- factor(group)
+    }
+
+    if (!is.null(label_threshold)) {
+      plot_data$labels <- ifelse(
+        abs(plot_data$elpd_diff) > label_threshold,
+        as.character(labels),
+        ""
+      )
+    }
+
+    jitter_position <- ggplot2::position_jitter(
+      width = jitter,
+      height = 0,
+      seed = 1
+    )
+
+    plot <- ggplot2::ggplot(
+      data = plot_data,
+      mapping = ggplot2::aes(x = y, y = elpd_diff)
+    ) +
+      ggplot2::geom_hline(yintercept = 0) +
+      ggplot2::labs(
+        x = if (sort_by_group) "Index" else NULL,
+        y = "Pointwise ELPD Difference (loo_1 - loo_2)"
+      )
+
+    if (is.null(group)) {
+      plot <- plot +
+        ggplot2::geom_point(
+          position = jitter_position,
+          alpha = alpha,
+          size = size
+        )
+    } else {
+      plot <- plot +
+        ggplot2::geom_point(
+          ggplot2::aes(color = group),
+          position = jitter_position,
+          alpha = alpha,
+          size = size
+        ) +
+        ggplot2::labs(color = "Group")
+    }
+
+    if (!is.null(label_threshold)) {
+      plot <- plot +
+        ggplot2::geom_text(
+          ggplot2::aes(label = labels),
+          position = jitter_position,
+          vjust = -0.5
+        )
+    }
+
+    plot
+  }
diff --git a/_pkgdown.yml b/_pkgdown.yml
index 0a216a02..9acc5235 100644
--- a/_pkgdown.yml
+++ b/_pkgdown.yml
@@ -95,6 +95,7 @@ reference:
       - loo_model_weights
       - stacking_weights
       - pseudobma_weights
+      - plot_loo_difference
   - title: Helper functions for K-fold CV
     contents:
       - kfold_split_random
diff --git a/man/plot_loo_difference.Rd b/man/plot_loo_difference.Rd
new file mode 100644
index 00000000..e628913c
--- /dev/null
+++ b/man/plot_loo_difference.Rd
@@ -0,0 +1,100 @@
+% Generated by roxygen2: do not edit by hand
+% Please edit documentation in R/loo_difference_plot.R
+\name{plot_loo_difference}
+\alias{plot_loo_difference}
+\title{Compare models across domains}
+\usage{
+plot_loo_difference(
+  y,
+  loo_1,
+  loo_2,
+  group = NULL,
+  size = 1,
+  alpha = 1,
+  jitter = 0,
+  sort_by_group = FALSE,
+  label_threshold = NULL,
+  labels = NULL
+)
+}
+\arguments{
+\item{y}{A vector of observations.}
+
+\item{loo_1, loo_2}{Objects returned by \code{\link[=loo]{loo()}}.}
+
+\item{group}{An optional grouping variable with the same length as \code{y}.
+Points are colored according to group membership.}
+
+\item{size, alpha}{Point size and opacity passed to \code{\link[ggplot2:geom_point]{ggplot2::geom_point()}}.}
+
+\item{jitter}{Amount of horizontal jitter passed to
+\code{\link[ggplot2:position_jitter]{ggplot2::position_jitter()}}.}
+
+\item{sort_by_group}{If \code{TRUE}, observations are ordered by \code{group}
+and the x-axis is replaced by a sequential index. The supplied \code{y} values
+are therefore not used as x coordinates. Plotting by index can be useful
+when categories have very different sample sizes. To control the group
+order, supply \code{group} as a factor with levels in the desired order.}
+
+\item{label_threshold}{Optional nonnegative threshold for labeling
+observations. Observations for which the absolute pointwise ELPD
+difference exceeds this value are labeled. If \code{NULL}, no observations
+are labeled.}
+
+\item{labels}{Optional vector of labels with the same length as \code{y}, used
+for observations selected by \code{label_threshold}. If \code{NULL}, observation
+indices are used.}
+}
+\value{
+A \code{\link[ggplot2:ggplot]{ggplot2::ggplot()}} object.
+}
+\description{
+The LOO difference plot shows how the ELPD of two different models
+changes when a predictor is varied. This can be useful for identifying
+opportunities for model stacking or expansion. Pointwise differences
+are computed as \code{loo_1 - loo_2}, so positive values indicate better
+predictive performance for \code{loo_1}.
+}
+\examples{
+# Artificial example
+log_lik <- example_loglik_matrix()
+shift <- seq(-0.5, 0.5, length.out = ncol(log_lik))
+log_lik_2 <- sweep(log_lik, 2, shift, FUN = "+")
+
+loo_1 <- loo(log_lik)
+loo_2 <- loo(log_lik_2)
+
+plot_loo_difference(
+  seq_len(ncol(log_lik)),
+  loo_1,
+  loo_2
+)
+
+# Label observations with large pointwise ELPD differences
+plot_loo_difference(
+  seq_len(ncol(log_lik)),
+  loo_1,
+  loo_2,
+  label_threshold = 0.3
+)
+
+# Create interspersed groups, then sort them in the plot
+group <- rep(c("A", "A", "A", "B"), length.out = ncol(log_lik))
+
+plot_loo_difference(
+  seq_len(ncol(log_lik)),
+  loo_1,
+  loo_2,
+  group = group,
+  sort_by_group = TRUE
+)
+
+}
+\references{
+Gabry, J. , Simpson, D. , Vehtari, A. , Betancourt, M. and
+Gelman, A. (2019), Visualization in Bayesian workflow.
+\emph{J. R. Stat. Soc. A}, 182: 389-402. doi:10.1111/rssa.12378
+(\href{https://rss.onlinelibrary.wiley.com/doi/full/10.1111/rssa.12378}{journal version},
+\href{https://arxiv.org/abs/1709.01449}{preprint arXiv:1709.01449},
+\href{https://github.com/jgabry/bayes-vis-paper}{code on GitHub})
+}
diff --git a/tests/testthat/test-loo_difference_plot.R b/tests/testthat/test-loo_difference_plot.R
new file mode 100644
index 00000000..7e294f55
--- /dev/null
+++ b/tests/testthat/test-loo_difference_plot.R
@@ -0,0 +1,125 @@
+skip_if_not_installed("ggplot2")
+
+log_lik <- example_loglik_matrix()
+shift <- seq(-0.5, 0.5, length.out = ncol(log_lik))
+log_lik_2 <- sweep(log_lik, 2, shift, FUN = "+")
+
+loo_1 <- loo(log_lik)
+loo_2 <- loo(log_lik_2)
+y <- seq_len(ncol(log_lik))
+
+test_that("plot_loo_difference returns pointwise ELPD differences", {
+  p <- plot_loo_difference(y, loo_1, loo_2)
+
+  expect_s3_class(p, "ggplot")
+  expect_equal(p$data$y, y)
+  expect_equal(p$data$elpd_diff, -shift)
+})
+
+test_that("plot_loo_difference sorts observations and labels by group", {
+  group <- factor(
+    rep(c("A", "B"), length.out = length(y)),
+    levels = c("B", "A")
+  )
+  labels <- paste0("obs", y)
+  ordering <- order(group)
+
+  p <- plot_loo_difference(
+    y,
+    loo_1,
+    loo_2,
+    group = group,
+    sort_by_group = TRUE,
+    label_threshold = 0.3,
+    labels = labels
+  )
+
+  expect_equal(p$data$y, seq_along(y))
+  expect_equal(p$data$elpd_diff, -shift[ordering])
+  expect_equal(
+    as.character(p$data$group),
+    as.character(group[ordering])
+  )
+  expect_equal(
+    p$data$labels,
+    ifelse(
+      abs(shift[ordering]) > 0.3,
+      labels[ordering],
+      ""
+    )
+  )
+})
+
+test_that("plot_loo_difference uses observation indices as default labels", {
+  p <- plot_loo_difference(
+    y,
+    loo_1,
+    loo_2,
+    label_threshold = 0.3
+  )
+
+  expect_equal(
+    p$data$labels,
+    ifelse(
+      abs(shift) > 0.3,
+      as.character(seq_along(y)),
+      ""
+    )
+  )
+})
+
+test_that("plot_loo_difference checks observation-level arguments", {
+  expect_error(
+    plot_loo_difference(y[-1], loo_1, loo_2)
+  )
+
+  expect_error(
+    plot_loo_difference(
+      y,
+      loo_1,
+      loo_2,
+      group = rep("A", length(y) - 1)
+    )
+  )
+
+  expect_error(
+    plot_loo_difference(
+      y,
+      loo_1,
+      loo_2,
+      label_threshold = -1
+    )
+  )
+
+  expect_error(
+    plot_loo_difference(
+      y,
+      loo_1,
+      loo_2,
+      label_threshold = 0.3,
+      labels = y[-1]
+    )
+  )
+
+  expect_error(
+    plot_loo_difference(
+      y,
+      loo_1,
+      loo_2,
+      labels = y
+    ),
+    "`label_threshold` must be supplied when `labels` is supplied.",
+    fixed = TRUE
+  )
+
+  expect_error(
+    plot_loo_difference(
+      y,
+      loo_1,
+      loo_2,
+      sort_by_group = TRUE
+    ),
+    "`group` must be supplied",
+    fixed = TRUE
+  )
+})