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

[Copilot]

For models predicting a value per response category (categorical, ordinal/cumulative, multinomial, or multivariate outcomes), estimate_contrasts(backend = "marginaleffects") folded the response category label into the Level1/Level2 strings (e.g. Level1 = "setosa 3"), confounding the contrasted factor level with the response category and making results hard to read/filter.

# Before
#> Level1 | Level2 | Difference
#> setosa 3 | setosa 2 | ...

# After
#> Level1 | Response1 | Level2 | Response2 | Difference
#> 3      | setosa    | 2      | setosa    | ...

Changes

• R/format.R
  • Added .split_contrast_response_levels(): detects response-category models (via cached model_info attribute, falling back to insight::model_info()/insight::is_multivariate()), and splits the merged Level1/Level2 strings into new Response1/Response2 columns.
  • Added .split_response_prefix(): regex-based helper that strips the known response-level prefix from each merged string.
  • Output column order is Level1, Response1, Level2, Response2, .... Response and level factors preserve the model's original response-level ordering.
  • The split is only applied when it unambiguously succeeds for every row; otherwise the original (unsplit) output is preserved, so ordinary contrasts and non-response-category models are unaffected.

Tests

• Added a regression test using nnet::multinom() verifying the new column names/order and cross-checking split contrasts against estimate_means() differences for pairing correctness.

[Copilot]

Pull request overview

This PR improves the readability and downstream usability of estimate_contrasts(backend = "marginaleffects") for response-category models (e.g., categorical/ordinal/multinomial and multivariate outcomes) by splitting response-category labels out of the contrasted level strings. Instead of embedding the response category inside Level1/Level2, the formatted output can include separate Response1/Response2 columns.

Changes:

• Add response/level splitting logic to the marginaleffects contrasts formatter so Level1/Level2 no longer confound response category labels.
• Add a regression test for nnet::multinom() verifying new columns/order and contrast correctness against estimate_means() deltas.
• Document the user-visible output change in NEWS.md and bump the development version.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

File	Description
R/format.R	Splits merged response-category prefixes out of contrast level strings and injects Response1/Response2 into formatted contrast outputs.
tests/testthat/test-estimate_contrasts.R	Adds regression coverage for multinomial response models to ensure split columns and correct pairing.
NEWS.md	Notes the new Response1/Response2 columns for response-category models.
DESCRIPTION	Bumps development version to reflect a user-visible change.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[strengejacke]

@DominiqueMakowski Can you check if this is the intended output?

library(modelbased)
data(iris)
m <- nnet::multinom(Species ~ Sepal.Width, data = iris, trace = FALSE)

out <- estimate_contrasts(m, contrast = "Sepal.Width = c(2, 3)")
print(out, table_width = Inf)
#> Marginal Contrasts Analysis
#> 
#> Level1 | Response1  | Level2 | Response2  | Difference |   SE |         95% CI | t(146) |      p
#> ------------------------------------------------------------------------------------------------
#> 3      | setosa     | 2      | setosa     |       0.21 | 0.04 | [ 0.12,  0.29] |   4.79 | < .001
#> 2      | versicolor | 2      | setosa     |       0.85 | 0.08 | [ 0.70,  1.01] |  10.68 | < .001
#> 3      | versicolor | 2      | setosa     |       0.34 | 0.05 | [ 0.25,  0.44] |   7.06 | < .001
#> 2      | virginica  | 2      | setosa     |       0.14 | 0.08 | [-0.01,  0.30] |   1.82 |  0.072
#> 3      | virginica  | 2      | setosa     |       0.45 | 0.05 | [ 0.34,  0.55] |   8.74 | < .001
#> 2      | versicolor | 3      | setosa     |       0.65 | 0.09 | [ 0.47,  0.82] |   7.20 | < .001
#> 3      | versicolor | 3      | setosa     |       0.14 | 0.08 | [-0.02,  0.29] |   1.75 |  0.083
#> 2      | virginica  | 3      | setosa     |      -0.06 | 0.09 | [-0.25,  0.12] |  -0.70 |  0.483
#> 3      | virginica  | 3      | setosa     |       0.24 | 0.08 | [ 0.08,  0.40] |   2.92 |  0.004
#> 3      | versicolor | 2      | versicolor |      -0.51 | 0.10 | [-0.70, -0.32] |  -5.33 | < .001
#> 2      | virginica  | 2      | versicolor |      -0.71 | 0.16 | [-1.02, -0.40] |  -4.48 | < .001
#> 3      | virginica  | 2      | versicolor |      -0.41 | 0.09 | [-0.59, -0.23] |  -4.41 | < .001
#> 2      | virginica  | 3      | versicolor |      -0.20 | 0.09 | [-0.38, -0.02] |  -2.23 |  0.027
#> 3      | virginica  | 3      | versicolor |       0.10 | 0.09 | [-0.07,  0.28] |   1.15 |  0.252
#> 3      | virginica  | 2      | virginica  |       0.30 | 0.09 | [ 0.11,  0.49] |   3.18 |  0.002
#> 
#> Variable predicted: Species
#> Predictors contrasted: Sepal.Width = c(2, 3)
#> p-values are uncorrected.
#> Contrasts are on the probs-scale.

estimate_means(m, by = "Sepal.Width = c(2, 3)")
#> Estimated Marginal Means
#> 
#> Sepal.Width | Response   | Probability |       SE |        95% CI | t(146)
#> --------------------------------------------------------------------------
#> 2           | setosa     |    1.14e-03 | 1.22e-03 | [ 0.00, 0.00] |   0.93
#> 2           | versicolor |        0.85 |     0.08 | [ 0.70, 1.01] |  10.75
#> 2           | virginica  |        0.14 |     0.08 | [-0.01, 0.30] |   1.82
#> 3           | setosa     |        0.21 |     0.04 | [ 0.12, 0.30] |   4.73
#> 3           | versicolor |        0.34 |     0.05 | [ 0.25, 0.44] |   7.11
#> 3           | virginica  |        0.45 |     0.05 | [ 0.35, 0.55] |   8.86
#> 
#> Variable predicted: Species
#> Predictors modulated: Sepal.Width = c(2, 3)
#> Predictions are on the probs-scale.

^{Created on 2026-07-20 with reprex v2.1.1}

[DominiqueMakowski]

yes, the output is now more readable and clear, and it is straightforward to manually filter the output as desired