[jmgirard]

For non-linear models, the footer of link-scale output printed "FALSE-transformed link-scale" or "TRUE-transformed link-scale". .table_footer() (R/table_footer.R) pasted transform into the text. But most callers store transform as TRUE or FALSE, and the emmeans code always stores TRUE.

Example, with d <- mtcars; d$cyl <- factor(d$cyl) and g <- glm(am ~ wt + cyl, data = d, family = binomial):

Call	Before	Now
estimate_link(g, by = "cyl")	FALSE-transformed link-scale	link-scale
estimate_means(g, "cyl", predict = "link", backend = "emmeans")	TRUE-transformed link-scale	link-scale
estimate_means(g, "cyl", predict = "link", transform = exp)	TRUE-transformed link-scale	transformed link-scale
estimate_contrasts(g, "cyl", predict = "link", transform = exp)	error: cannot coerce type 'builtin' to vector of type 'character'	transformed link-scale
estimate_contrasts(g, "cyl", predict = "link", transform = "exp")	exp-transformed link-scale	exp-transformed link-scale
estimate_expectation(g, data = estimate_link(g, by = "cyl"))	FALSE-transformed link-scale	response-scale

What changed:

• A new attribute transform_arg keeps the transform argument as the user passed it. get_marginalmeans(), get_marginalcontrasts() and .estimate_predicted() set it. The new helper .link_transform_label() reads it for the link-scale line. A name such as "exp" gives "exp-transformed", and a function gives "transformed". If insight::find_transformation() finds a response transformation, transform = TRUE gives "transformed". Otherwise it gives "link-scale", because nothing changes. The logical transform attribute stays as it was, because the plot code and the linear-model hint read it.
• An earlier output passed as data brings its own predict and transform attributes into .estimate_predicted(). Before, c(grid_specs, list(predict = predict), ...) put them first, so info$predict took the old value. Now the values of the call replace them.

Tests in test-table_footer.R cover the calls above with both backends, and transform = TRUE. They also cover an earlier output as data in both directions, contrasts of estimate_link() output, and contrasts of slopes.

Not fixed here:

• On the response scale, the footer does not say that a transform function changed the values.
• estimate_link(g, transform = "exp") with a string stops with "could not find function trans_fun".
• backend = "emmeans" ignores a transform function in estimate_means().
• With an earlier output of another model as data, the footer reads the model_info of that output.

devtools::test() gives 0 failures and 0 errors locally. devtools::check() gives 0 errors and 0 warnings. The version is bumped to 0.17.0.17.

[Copilot]

🟡 Changes recommended

Add coverage for transformed-response footer labeling and correct the remaining comment typo.

1 open issue

• Correct misspelling of response in comment · New

What changed in this PR

Fixes link-scale footer labels and preserves user-supplied transformations across prediction and contrast workflows.

Changes:

• Adds transformation-aware footer labeling.
• Preserves transformation metadata and corrects precedence when reusing outputs.
• Adds regression tests and bumps the package version.

File	Summary
tests/​testthat/​test-table_footer.R	Adds footer regression coverage.
R/​table_footer.R	Generates corrected link-scale labels; transformed-response coverage is still needed.
R/​get_marginalmeans.R	Stores the original transform argument; includes a comment typo.
R/​get_marginalcontrasts.R	Propagates transformation metadata.
R/​estimate_predicted.R	Corrects inherited metadata precedence.
NEWS.md	Documents the fixes.
DESCRIPTION	Bumps the development version.

_{Review effort: Lite}