Contrasting slopes: differences between backends

[DominiqueMakowski]

emmeans acts up funny:

model <- lm(Petal.Width ~ Petal.Length * Species, data = iris)

modelbased::estimate_contrasts(model, contrast="Petal.Length", by="Species", backend="marginaleffectss")
#> Marginal Contrasts Analysis
#> 
#> Parameter              | Coefficient |   SE |        95% CI |     t |     p
#> ---------------------------------------------------------------------------
#> setosa - versicolor    |       -0.13 | 0.16 | [-0.43, 0.17] | -0.83 | 0.808
#> setosa - virginica     |        0.04 | 0.15 | [-0.26, 0.34] |  0.27 | 0.808
#> versicolor - virginica |        0.17 | 0.07 | [ 0.03, 0.31] |  2.41 | 0.048
#> 
#> Marginal contrasts estimated at Petal.Length
#> p-value adjustment method: Holm (1979)
modelbased::estimate_contrasts(model, contrast="Petal.Length", by="Species", backend="emmeans")
#> Marginal Contrasts Analysis
#> 
#> Level1                       |                       Level2 |    Species
#> ------------------------------------------------------------------------
#> Petal.Length1                | Petal.Length1.65555555555556 |     setosa
#> Petal.Length1                | Petal.Length1.65555555555556 | versicolor
#> Petal.Length1                | Petal.Length1.65555555555556 |  virginica
#> Petal.Length1                | Petal.Length2.31111111111111 |     setosa
#> Petal.Length1                | Petal.Length2.31111111111111 | versicolor
#> Petal.Length1                | Petal.Length2.31111111111111 |  virginica
#> Petal.Length1                | Petal.Length2.96666666666667 |     setosa
#> Petal.Length1                | Petal.Length2.96666666666667 | versicolor
#> Petal.Length1                | Petal.Length2.96666666666667 |  virginica
#> Petal.Length1                | Petal.Length3.62222222222222 |     setosa
#> Petal.Length1                | Petal.Length3.62222222222222 | versicolor
#> Petal.Length1                | Petal.Length3.62222222222222 |  virginica
#> ...
#> 
#> Marginal contrasts estimated at Petal.Length
#> p-value adjustment method: Holm (1979)

^{Created on 2025-01-13 with reprex v2.1.1}

[strengejacke]

You're a bit out of date with the modelbased version, the first example currently doesn't work. Fixed in #331

[strengejacke]

The question is, how much work do we want to put in maintaining the emmeans backend? I haven't looked much into the code for emmeans yet.

[DominiqueMakowski]

Not sure long term, but short term it'd be good to be able to show that we can reproduce its result as - at least in my mind - emmeans is kind of the baseline / ground truth.

[strengejacke]

Ok, and we want:

model <- lm(Petal.Width ~ Petal.Length * Species, data = iris)
emmeans::emtrends(model, "Species", "Petal.Length") |>
  emmeans::contrast(method = "pairwise")
#>  contrast               estimate     SE  df t.ratio p.value
#>  setosa - versicolor     -0.1298 0.1560 144  -0.835  0.6821
#>  setosa - virginica       0.0409 0.1530 144   0.268  0.9613
#>  versicolor - virginica   0.1708 0.0708 144   2.412  0.0449
#> 
#> P value adjustment: tukey method for comparing a family of 3 estimates

^{Created on 2025-02-13 with reprex v2.1.1}