estimate_slopes() with custom contrasts works with marginaleffects but not emmeans

[betahat]

If I use the example from the documentation:

data(puppy_love, package = "modelbased")
cond_tx <- cbind("no treatment" = c(1, 0, 0), "treatment" = c(0, 0.5, 0.5))
model <- lm(happiness ~ puppy_love * dose, data = puppy_love)
estimate_slopes(model, "puppy_love", by = "dose", comparison = cond_tx)

things look fine, but if I specify the backend:

estimate_slopes(model, "puppy_love", by = "dose", comparison = cond_tx, backend = "emmeans")
estimate_slopes(model, "puppy_love", by = "dose", comparison = cond_tx, backend = "marginaleffects")

The emmeans gives output of:

Estimated Marginal Effects

dose       | Slope |   SE |        95% CI | t(24) |     p
---------------------------------------------------------
No puppies |  0.76 | 0.27 | [ 0.21, 1.31] |  2.86 | 0.009
15 mins    |  0.82 | 0.34 | [ 0.11, 1.53] |  2.40 | 0.025
30 mins    | -0.22 | 0.28 | [-0.79, 0.35] | -0.79 | 0.436

where the marginaleffects one gives output for the custom contrasts:

Estimated Marginal Effects

Parameter    | Slope |   SE |        95% CI | t(24) |     p
-----------------------------------------------------------
no treatment |  0.76 | 0.27 | [ 0.21, 1.31] |  2.86 | 0.009
treatment    |  0.30 | 0.22 | [-0.15, 0.75] |  1.37 | 0.184

Marginal effects estimated for puppy_love

[strengejacke]

marginaleffects can calculate comparisons / contrasts for the different functions, not only for avg_comparisons(). Thus, you can simply "force" other functions to calculate contrasts / comparisons.

In emmeans, you have to use the contrast function explicitly. Here's how to make your example work:

library(modelbased)
data(puppy_love, package = "modelbased")
model <- lm(happiness ~ puppy_love * dose, data = puppy_love)

cond_tx <- list("no treatment" = c(1, 0, 0), "treatment" = c(0, 0.5, 0.5))
estimate_contrasts(model, "puppy_love", by = "dose", comparison = cond_tx, backend = "emmeans")
#> Marginal Contrasts Analysis
#> 
#> Level        | Difference |        95% CI |   SE | t(24) |     p
#> ----------------------------------------------------------------
#> no treatment |       0.76 | [ 0.21, 1.31] | 0.27 |  2.86 | 0.009
#> treatment    |       0.30 | [-0.15, 0.75] | 0.22 |  1.37 | 0.184
#> 
#> Variable predicted: happiness
#> Predictors contrasted: puppy_love
#> p-values are uncorrected.

cond_tx <- cbind("no treatment" = c(1, 0, 0), "treatment" = c(0, 0.5, 0.5))
estimate_slopes(
  model,
  "puppy_love",
  by = "dose",
  comparison = cond_tx,
  backend = "marginaleffects"
)
#> Estimated Marginal Effects
#> 
#> Parameter    | Slope |   SE |        95% CI | t(24) |     p
#> -----------------------------------------------------------
#> no treatment |  0.76 | 0.27 | [ 0.21, 1.31] |  2.86 | 0.009
#> treatment    |  0.30 | 0.22 | [-0.15, 0.75] |  1.37 | 0.184
#> 
#> Marginal effects estimated for puppy_love

^{Created on 2026-05-16 with reprex v2.1.1}

[strengejacke]

We could check if a comparison argument is used, and then silently call estimate_contrasts() when the backend is set to "emmeans".

[strengejacke]

We could check if a comparison argument is used, and then silently call estimate_contrasts() when the backend is set to "emmeans".

Just did a quick check, I think this will lower code readability and make it harder to maintain, so we must think about if/how we can implement this.