strengejacke
commented
3 years ago Not sure, but I think that emtrends calculates marginal effects. Maybe you can add an example that compares marginaleffects to modelbased to see if results match? Here's a comparison:
# marginal effect
model <- lm(Sepal.Width ~ Species * Petal.Length, data = iris)
# own class/print
modelbased::estimate_slopes(model, trend = "Petal.Length", at = "Species")
#> Estimated Marginal Effects
#>
#> Species | Coefficient | SE | 95% CI | t(144) | p
#> -----------------------------------------------------------------
#> setosa | 0.39 | 0.26 | [-0.13, 0.90] | 1.49 | 0.138
#> versicolor | 0.37 | 0.10 | [ 0.18, 0.56] | 3.89 | < .001
#> virginica | 0.23 | 0.08 | [ 0.07, 0.40] | 2.86 | 0.005
#> Marginal effects estimated for Petal.Length
# emtrends, with modelbased API
modelbased::model_emtrends(model, trend = "Petal.Length", at = "Species")
#> Species Petal.Length.trend SE df lower.CL upper.CL
#> setosa 0.388 0.2602 144 -0.1264 0.902
#> versicolor 0.374 0.0961 144 0.1843 0.564
#> virginica 0.234 0.0819 144 0.0725 0.396
#>
#> Confidence level used: 0.95
# emmeans::emtrends
emmeans::emtrends(model, c("Petal.Length", "Species"), "Petal.Length")
#> Petal.Length Species Petal.Length.trend SE df lower.CL upper.CL
#> 3.76 setosa 0.388 0.2602 144 -0.1264 0.902
#> 3.76 versicolor 0.374 0.0961 144 0.1843 0.564
#> 3.76 virginica 0.234 0.0819 144 0.0725 0.396
#>
#> Confidence level used: 0.95
# marginal mean
model <- lm(Sepal.Width ~ Species + Petal.Length, data = iris)
modelbased::estimate_means(model, at = "Petal.Length=c(1,4,6)")
#> Estimated Marginal Means
#>
#> Petal.Length | Mean | SE | 95% CI
#> -----------------------------------------
#> 1.00 | 2.23 | 0.17 | [1.90, 2.57]
#> 4.00 | 3.13 | 0.03 | [3.07, 3.19]
#> 6.00 | 3.73 | 0.14 | [3.45, 4.00]
#>
#> Marginal means estimated at Petal.Length
modelbased::model_emmeans(model, at = "Petal.Length=c(1,4,6)")
#> Petal.Length emmean SE df lower.CL upper.CL
#> 1 2.23 0.1688 146 1.90 2.57
#> 4 3.13 0.0296 146 3.07 3.19
#> 6 3.73 0.1380 146 3.45 4.00
#>
#> Results are averaged over the levels of: Species
#> Confidence level used: 0.95
emmeans::emmeans(model, "Petal.Length", at = list(Petal.Length = c(1, 4, 6)))
#> Petal.Length emmean SE df lower.CL upper.CL
#> 1 2.23 0.1688 146 1.90 2.57
#> 4 3.13 0.0296 146 3.07 3.19
#> 6 3.73 0.1380 146 3.45 4.00
#>
#> Results are averaged over the levels of: Species
#> Confidence level used: 0.95
# adjusted predictions
model <- lm(Sepal.Width ~ Species + Petal.Length, data = iris)
vm <- modelbased::visualisation_matrix(iris, at = "Petal.Length=c(1,4,6)")
modelbased::estimate_expectation(model, vm)
#> Model-based Expectation
#>
#> Petal.Length | Species | Predicted | SE | 95% CI | Residuals
#> --------------------------------------------------------------------
#> 1.00 | setosa | 3.29 | 0.05 | [3.19, 3.39] | 0.23
#> 4.00 | setosa | 4.19 | 0.16 | [3.87, 4.50] | 1.13
#> 6.00 | setosa | 4.78 | 0.28 | [4.23, 5.33] | 1.72
#>
#> Variable predicted: Sepal.Width
#> Predictors modulated: Petal.Length=c(1,4,6)
#> Predictors controlled: Sepal.Length
predict(model, newdata = vm, interval = "confidence")
#> fit lwr upr
#> 1 3.290180 3.186116 3.394244
#> 2 4.185113 3.869171 4.501055
#> 3 4.781735 4.232160 5.331310Created on 2021-11-07 by the reprex package (v2.0.1)
vincentarelbundock
bwiernik
DominiqueMakowski
mattansb
Hi all,
As most of you know, I recently released {marginaleffects}, a package with functionality that overlaps and (I think/hope) complements {modelbased} and {ggeffects}:
https://vincentarelbundock.github.io/marginaleffects/
Definitions
There seems to be a ton of variation in how vocabulary is used in this space, so here is how I define the 4 quantities that {marginaleffects} can compute:
To produce all four of these quantities, the {marginaleffects} package requires a model object (mandatory) and a dataset (optional).
What does {modelbased} need?
I admit that I have very little experience with {modelbased}, so I wanted to ask you all: Is there anything in the list above that would be of interest here? Or do you already cover all this territory?
From a cursory review, my sense is that Adjusted Predictions are already very well-covered. There is probably nothing I can contribute here.
Marginal Means also seem well-covered. Are you doing the computation yourself, or wrapping
emmeans::emmeans? Are you happy with this, or is there anything missing?{modelbased} can also compute pairwise contrasts. Is that also by wrapping
emmeans? In Issue https://github.com/easystats/modelbased/issues/150, some people expressed an interest in expanding the range of contrasts supported. This could be done by enriching theemmeanswrapper. Currently, {marginaleffects} only supports pairwise contrasts, but the infrastructure would make it relatively easy to allow users to specify specific contrasts. That said, I am unlikely to reach parity withemmeansin the medium run.Finally, does {modelbased} support "marginal effects" as defined above (i.e., slopes / partial derivatives)? The website suggests that the answer is "yes", but I admit that I am a bit confused by the nomenclature, especially this section of your docs:
Your concept of "marginal effect" does not sound like the same as mine... In my definition, the partial derivative (slope) of the regression equation is the continuous analogue to a contrast for categorical predictors. Marginal effects measure the change inY that is associated with an infinitesimal change in a continuous X. A contrast measures the change in Y that is associated with a level change in a categorical X. But perhaps this is just nomenclature, and you already support this.
What does the future hold?
I opened this issue to know if there was anything I could bring to the table. My sense is that the slope computation is the most "mature" and useful aspect of my package, so I would guess that this is where most of the value lies. But if you feel that {modelbased} already covers the 4 quantities defined above adequately, then the answer is: "No, there is no added value." That's fine, and would be useful for me to know.
On the other hand, if there are some features that you would like to add but are having trouble with, I would be very curious to hear about it. Some of them might be easy to add.