diff --git a/NEWS.md b/NEWS.md
index b69d472f44..ff33dc03d2 100644
--- a/NEWS.md
+++ b/NEWS.md
@@ -1,5 +1,7 @@
 # ggplot2 (development version)
 
+* Extended `stat_ecdf()` to calculate the cdf from either x or y instead from y only (@jgjl, #4005).
+
 * Fixed a bug in `labeller()` so that `.default` is passed to `as_labeller()`
   when labellers are specified by naming faceting variables. (@waltersom, #4031)
   
diff --git a/R/stat-ecdf.r b/R/stat-ecdf.r
index a974719dde..8f6654c67e 100644
--- a/R/stat-ecdf.r
+++ b/R/stat-ecdf.r
@@ -7,6 +7,11 @@
 #' The downside is that it requires more training to accurately interpret,
 #' and the underlying visual tasks are somewhat more challenging.
 #'
+#' The statistic relies on the aesthetics assignment to guess which variable to
+#' use as the input and which to use as the output. Either x or y must be provided
+#' and one of them must be unused. The ECDF will be calculated on the given aesthetic
+#' and will be output on the unused one.
+#'
 #' @inheritParams layer
 #' @inheritParams geom_point
 #' @param na.rm If `FALSE` (the default), removes missing values with
@@ -17,7 +22,6 @@
 #'   and (Inf, 1)
 #' @section Computed variables:
 #' \describe{
-#'   \item{x}{x in data}
 #'   \item{y}{cumulative density corresponding x}
 #' }
 #' @export
@@ -67,7 +71,24 @@ stat_ecdf <- function(mapping = NULL, data = NULL,
 #' @usage NULL
 #' @export
 StatEcdf <- ggproto("StatEcdf", Stat,
-  compute_group = function(data, scales, n = NULL, pad = TRUE) {
+  required_aes = c("x|y"),
+
+  default_aes = aes(y = after_stat(y)),
+
+  setup_params = function(data, params) {
+    params$flipped_aes <- has_flipped_aes(data, params, main_is_orthogonal = FALSE, main_is_continuous = TRUE)
+
+    has_x <- !(is.null(data$x) && is.null(params$x))
+    has_y <- !(is.null(data$y) && is.null(params$y))
+    if (!has_x && !has_y) {
+      abort("stat_ecdf() requires an x or y aesthetic.")
+    }
+
+    params
+  },
+
+  compute_group = function(data, scales, n = NULL, pad = TRUE, flipped_aes = FALSE) {
+    data <- flip_data(data, flipped_aes)
     # If n is NULL, use raw values; otherwise interpolate
     if (is.null(n)) {
       x <- unique(data$x)
@@ -78,13 +99,11 @@ StatEcdf <- ggproto("StatEcdf", Stat,
     if (pad) {
       x <- c(-Inf, x, Inf)
     }
-    y <- ecdf(data$x)(x)
-
-    new_data_frame(list(x = x, y = y), n = length(x))
-  },
-
-  default_aes = aes(y = after_stat(y)),
+    data_ecdf <- ecdf(data$x)(x)
 
-  required_aes = c("x")
+    df_ecdf <- new_data_frame(list(x = x, y = data_ecdf), n = length(x))
+    df_ecdf$flipped_aes <- flipped_aes
+    flip_data(df_ecdf, flipped_aes)
+  }
 )
 
diff --git a/man/stat_ecdf.Rd b/man/stat_ecdf.Rd
index 12afbfb07e..58423c3d41 100644
--- a/man/stat_ecdf.Rd
+++ b/man/stat_ecdf.Rd
@@ -76,10 +76,15 @@ tuning parameters and handles both continuous and categorical variables.
 The downside is that it requires more training to accurately interpret,
 and the underlying visual tasks are somewhat more challenging.
 }
+\details{
+The statistic relies on the aesthetics assignment to guess which variable to
+use as the input and which to use as the output. Either x or y must be provided
+and one of them must be unused. The ECDF will be calculated on the given aesthetic
+and will be output on the unused one.
+}
 \section{Computed variables}{
 
 \describe{
-\item{x}{x in data}
 \item{y}{cumulative density corresponding x}
 }
 }
diff --git a/tests/testthat/test-stat-ecdf.R b/tests/testthat/test-stat-ecdf.R
new file mode 100644
index 0000000000..4e56e3625b
--- /dev/null
+++ b/tests/testthat/test-stat-ecdf.R
@@ -0,0 +1,16 @@
+context("stat_ecdf")
+
+test_that("stat_ecdf works in both directions", {
+  p <- ggplot(mpg, aes(hwy)) + stat_ecdf()
+  x <- layer_data(p)
+  expect_false(x$flipped_aes[1])
+
+  p <- ggplot(mpg, aes(y = hwy)) + stat_ecdf()
+  y <- layer_data(p)
+  expect_true(y$flipped_aes[1])
+
+  x$flipped_aes <- NULL
+  y$flipped_aes <- NULL
+  expect_identical(x, flip_data(y, TRUE)[,names(x)])
+})
+