estimate_contrasts: split Response labels from Level1/Level2 for response-category models

DESCRIPTION

@@ -1,7 +1,7 @@
 Type: Package
 Package: modelbased
 Title: Estimation of Model-Based Predictions, Contrasts and Means
-Version: 0.16.0.10
+Version: 0.16.0.11
 Authors@R:
     c(person(given = "Dominique",
              family = "Makowski",

NEWS.md

@@ -8,6 +8,12 @@
 
 ## Changes
 
+* For contrasts of models where predictions vary by response category
+  (e.g., categorical, ordinal/cumulative or multinomial models, as well as
+  multivariate response models), `estimate_contrasts()` now returns separate
+  `Response1` and `Response2` columns instead of merging the response
+  category label into the `Level1` and `Level2` columns (#646).
+
 * The `iterations` argument can now also be used for Bayesian model in
   `estimate_means()`, `estimate_slopes()`, and `estimate_contrasts()`, which
   is then passed to the `ndraws` argument and controls how many samples are

R/format.R

@@ -512,9 +512,23 @@ format.marginaleffects_contrasts <- function(
         }
       }
 
+      # for models where predictions are made for each level of the response
+      # (e.g., categorical, ordinal/cumulative or multinomial models, as well
+      # as multivariate response models), marginaleffects merges the response
+      # level and the contrasted factor level into the same string (e.g.,
+      # "setosa 3"). We split those into separate "Response1" and "Response2"
+      # columns, so contrasted factor levels and response levels are not
+      # confounded
+      params <- .split_contrast_response_levels(params, model, x)
+
       # we need to update these variables, because these are the new column
-      # names for contrasts and focal terms
-      contrast <- c("Level1", "Level2")
+      # names for contrasts and focal terms. We use `intersect()` here because
+      # "Response1"/"Response2" are only present if the contrasted factor and
+      # response category levels were successfully split above
+      contrast <- intersect(
+        c("Level1", "Response1", "Level2", "Response2"),
+        colnames(params)
+      )
 
       # remove old column
       x$Parameter <- NULL
@@ -570,6 +584,95 @@ format.marginaleffects_contrasts <- function(
 # variable names to levels. We first find the intersection of all
 # levels from all current contrast terms
 
+# for models where predictions are made separately for each level of the
+# response (categorical, ordinal/cumulative, multinomial or multivariate
+# models), marginaleffects merges the response level and the contrasted
+# factor level into the same string, e.g. "setosa 3". This function splits
+# "Level1" and "Level2" into "Response1"/"Level1" and "Response2"/"Level2",
+# so response levels and contrasted factor levels are not confounded. If
+# splitting isn't possible (e.g., not a response-level model, or the response
+# levels cannot be unambiguously detected), `params` is returned unchanged.
+
+.split_contrast_response_levels <- function(params, model, x = NULL) {
+  if (is.null(model) || !all(c("Level1", "Level2") %in% colnames(params))) {
+    return(params)
+  }
+
+  # prefer the model info already stored as attribute (consistent with how
+  # other format-methods retrieve model information), and only compute it
+  # ourselves as a fallback
+  model_info <- attr(x, "model_info", exact = TRUE)
+  if (is.null(model_info)) {
+    model_info <- .safe(insight::model_info(model, verbose = FALSE))
+  }
+  is_multivariate <- isTRUE(.safe(insight::is_multivariate(model)))
+  is_response_model <- isTRUE(model_info$is_multinomial) ||
+    isTRUE(model_info$is_categorical) ||
+    isTRUE(model_info$is_ordinal) ||
+    isTRUE(model_info$is_cumulative) ||
+    is_multivariate
+
+  if (!is_response_model) {
+    return(params)
+  }
+
+  # response levels, used to detect and split off the response-category
+  # label from the contrasted factor level. We keep the original order of
+  # the response levels (e.g., factor level order), so that "Response1" and
+  # "Response2" use a meaningful, consistent level order
+  response <- .safe(insight::get_response(model, verbose = FALSE))
+  response_levels <- if (is.factor(response)) {
+    levels(response)
+  } else {
+    unique(as.character(response))
+  }
+  if (is_multivariate) {
+    response_levels <- unique(c(response_levels, insight::find_response(model)))
+  }
+  if (!length(response_levels)) {
+    return(params)
+  }
+
+  # sort by length, descending, to avoid partial matches for levels that are
+  # substrings of other levels (used for pattern matching only)
+  search_levels <- response_levels[order(nchar(response_levels), decreasing = TRUE)]
+
+  split1 <- .split_response_prefix(params$Level1, search_levels)
+  split2 <- .split_response_prefix(params$Level2, search_levels)
+
+  # sanity check - only split if we found a response label for *all* rows, in
+  # both columns, otherwise keep the original (unsplit) columns
+  if (all(!is.na(split1$Response)) && all(!is.na(split2$Response))) {
+    level_order <- unique(c(split1$Level, split2$Level))
+    response_order <- intersect(response_levels, unique(c(split1$Response, split2$Response)))
+    params$Level1 <- factor(split1$Level, levels = level_order)
+    params$Response1 <- factor(split1$Response, levels = response_order)
+    params$Level2 <- factor(split2$Level, levels = level_order)
+    params$Response2 <- factor(split2$Response, levels = response_order)
+  }
+
+  params
+}
+
+
+# split off a known "response level" prefix from a vector of strings, e.g.
+# "setosa 3" -> Response = "setosa", Level = "3"
+
+.split_response_prefix <- function(x, response_levels) {
+  x_chr <- as.character(x)
+  out <- data.frame(Response = NA_character_, Level = x_chr, stringsAsFactors = FALSE)
+  for (i in response_levels) {
+    pattern <- paste0("^", i, "[[:space:]]+")
+    matches <- is.na(out$Response) & grepl(pattern, x_chr)
+    if (any(matches)) {
+      out$Response[matches] <- i
+      out$Level[matches] <- insight::trim_ws(sub(pattern, "", x_chr[matches]))
+    }
+  }
+  out
+}
+
+
 .fix_duplicated_contrastlevels <- function(params, contrast_names) {
   multiple_levels <- Reduce(
     function(i, j) intersect(i, j),

tests/testthat/test-estimate_contrasts.R

@@ -1815,3 +1815,42 @@ test_that("estimate_contrast, p-adjust tukey works for contrasting slopes", {
   # here to avoid fragile tests while still ensuring close agreement.
   expect_equal(out1$p.value, out2$p, tolerance = 1e-2)
 })
+
+
+test_that("estimate_contrast, categorical/multinomial response models split off Response levels", {
+  skip_if_not_installed("nnet")
+
+  data(iris)
+  m <- nnet::multinom(Species ~ Sepal.Width, data = iris, trace = FALSE)
+  out <- estimate_contrasts(m, contrast = "Sepal.Width = c(2, 3)", backend = "marginaleffects")
+
+  # Level1/Response1/Level2/Response2 columns, in that order
+  expect_named(
+    out,
+    c(
+      "Level1", "Response1", "Level2", "Response2", "Difference", "SE",
+      "CI_low", "CI_high", "t", "df", "p"
+    )
+  )
+  # contrasted levels no longer contain the response category label
+  expect_true(all(as.character(out$Level1) %in% c("2", "3")))
+  expect_true(all(as.character(out$Level2) %in% c("2", "3")))
+  expect_true(all(as.character(out$Response1) %in% levels(iris$Species)))
+  expect_true(all(as.character(out$Response2) %in% levels(iris$Species)))
+
+  # contrasts within the same response category should match the difference
+  # between the corresponding marginal means for that category
+  means <- estimate_means(m, by = "Sepal.Width = c(2, 3)", backend = "marginaleffects")
+  for (resp in levels(iris$Species)) {
+    row <- out[
+      as.character(out$Level1) == "3" &
+        as.character(out$Level2) == "2" &
+        as.character(out$Response1) == resp &
+        as.character(out$Response2) == resp,
+    ]
+    expect_identical(nrow(row), 1L)
+    mean_3 <- means$Probability[means$Sepal.Width == "3" & means$Response == resp]
+    mean_2 <- means$Probability[means$Sepal.Width == "2" & means$Response == resp]
+    expect_equal(row$Difference, mean_3 - mean_2, tolerance = 1e-4)
+  }
+})