Full support for Bayesian models when backend = "marginaleffects"

[DominiqueMakowski]

This issue might be related to parameters(). Sometimes we get indices like pd, sometimes not.

df <- iris
df$Condition <- rep(c("A", "B", "C"), length.out=150)

m1 <- brms::brm(Sepal.Width ~ Species * Sepal.Length, data=df, refresh=0)


# pd is present
modelbased::estimate_slopes(m1, trend="Sepal.Length", by="Species")
#> Estimated Marginal Effects
#> 
#> Species    | Coefficient |       95% CI |     pd
#> ------------------------------------------------
#> setosa     |        0.80 | [0.57, 1.01] |   100%
#> versicolor |        0.32 | [0.17, 0.47] | 99.98%
#> virginica  |        0.23 | [0.11, 0.35] |   100%
#> Marginal effects estimated for Sepal.Length

# BUT, pd not present here
trends <- modelbased::get_marginaltrends(m1, trend="Sepal.Length", by="Species")
parameters::parameters(trends)
#> # Fixed Effects
#> 
#> Parameter    |    Species | Coefficient |       95% CI
#> ------------------------------------------------------
#> Sepal.Length |     setosa |        0.80 | [0.57, 1.01]
#> Sepal.Length | versicolor |        0.32 | [0.17, 0.47]
#> Sepal.Length |  virginica |        0.23 | [0.11, 0.35]


m2 <- brms::brm(Condition ~ Species, data=df, family=brms::categorical(refcat="A"), refresh=0)

# pd not present
modelbased::estimate_slopes(m2, trend="Species", backend="marginaleffects")
#> Estimated Marginal Effects
#> 
#> Response |          Comparison | Coefficient |        95% CI
#> ------------------------------------------------------------
#> A        | versicolor - setosa |    1.25e-03 | [-0.19, 0.18]
#> A        |  virginica - setosa |       -0.02 | [-0.20, 0.17]
#> B        | versicolor - setosa |       -0.02 | [-0.21, 0.16]
#> B        |  virginica - setosa |   -4.60e-04 | [-0.19, 0.19]
#> C        | versicolor - setosa |        0.02 | [-0.16, 0.20]
#> C        |  virginica - setosa |        0.02 | [-0.17, 0.20]
#> Marginal effects estimated for Species

^{Created on 2025-01-09 with reprex v2.1.1}

[strengejacke]

I think in the .format_marginaleffects_*() internals, we only call parameters::parameters(). For emmeans, we call bayestestR::describe_posterior() for Bayesian models, which we should also do for marginaleffects.

The implementation is not feature complete yet, we still have to put some work into the .format_marginaleffects_*() internals.

[DominiqueMakowski]

It could be done at parameters()' level though no? I would have expected parameters() to return these indices

[strengejacke]

Yes, for Bayesian models. But not for marginaleffects object yet (I think @vincentarelbundock implemented the marginaleffects methods in parameters, and currently, they're calling tidy(), which does not return those indices - in bayestestR, @mattansb implemented methods for marginaleffects objects that processed Bayesian models).

[mattansb]

Yes, I think Bayesian emmeans and Bayesian marginaleffects objects should be processes the same through bayestestR.

[DominiqueMakowski]

Bayesian emmeans and Bayesian marginaleffects objects should be processes the same through bayestestR.

Why not through parameters (which relies on bayestestR)?

[mattansb]

Should be, but is sounds like {parameters} is not using {bayestestR}?

[strengejacke]

Should be, but is sounds like {parameters} is not using {bayestestR}?

#302 (comment)