Overview
This package has functions to calculate marginal effects from brms
models ( http://paulbuerkner.github.io/brms/ ). A central motivator is to calculate average marginal effects (AMEs) for continuous and discrete predictors in fixed effects only and mixed effects regression models including location scale models.
This table shows an overview of currently supported models / features where “X” indicates a specific model / feature is currently supported. The column ‘Fixed’ means fixed effects only models. The column ‘Mixed’ means mixed effects models.
Distribution / Feature  Fixed  Mixed 

Gaussian / Normal  ✔️  ✔️ 
Bernoulli (logistic)  ✔️  ✔️ 
Poisson  ✔️  ✔️ 
Negative Binomial  ✔️  ✔️ 
Gamma  ✔️  ✔️ 
Beta  ✔️  ✔️ 
Multinomial logistic  ❌  ❌ 
Multivariate models  ❌  ❌ 
Gaussian location scale models  ✔️  ✔️ 
Natural log / square root transformed outcomes  ✔️  ✔️ 
Monotonic predictors  ✔️  ✔️ 
Custom outcome transformations  ❌  ❌ 
In general, any distribution supported by brms
that generates one and only one predicted value (e.g., not multinomial logistic regression models) should be supported for fixed effects only models. Also note that currently, only Gaussian random effects are supported. This is not too limiting as even for Bernoulli, Poisson, etc. outcomes, the random effects are commonly assumed to have a Gaussian distribution.
Here is a quick syntax overview of how to use the main function, brmsmargins()
.
Fixed effects, continuous predictor.
h < .001
ames < brmsmargins(
object = model,
add = data.frame(x = c(0, h)),
contrasts = cbind("AME x" = c(1 / h, 1 / h)),
effects = "fixedonly")
ames$ContrastSummary
Fixed effects, discrete predictor.
ames < brmsmargins(
object = model,
add = data.frame(x = c(0, 1)),
contrasts = cbind("AME x" = c(1, 1)),
effects = "fixedonly")
ames$Summary
ames$ContrastSummary
Mixed effects, continuous predictor.
h < .001
ames < brmsmargins(
object = model,
add = data.frame(x = c(0, h)),
contrasts = cbind("AME x" = c(1 / h, 1 / h)),
effects = "integrateoutRE")
ames$ContrastSummary
Mixed effects, discrete predictor.
ames < brmsmargins(
object = model,
add = data.frame(x = c(0, 1)),
contrasts = cbind("AME x" = c(1, 1)),
effects = "integrateoutRE")
ames$Summary
ames$ContrastSummary
Mixed Effects Location Scale, continuous predictor
h < .001
ames < brmsmargins(
object = model,
add = data.frame(x = c(0, h)),
contrasts = cbind("AME x" = c(1 / h, 1 / h)),
dpar = "sigma",
effects = "integrateoutRE")
ames$ContrastSummary
Mixed Effects Location Scale, discrete predictor
ames < brmsmargins(
object = model,
at = data.frame(x = c(0, 1)),
contrasts = cbind("AME x" = c(1, 1)),
dpar = "sigma",
effects = "integrateoutRE")
ames$Summary
ames$ContrastSummary
Note that even on mixed effects models, it is possible to generate predictions and marginal effects from the fixed effects only, just by specifying effects = "fixedonly"
but this is probably not a good idea generally so not shown by default.
Also note that for all of these examples ames$Summary
would have a summary of the averaged predicted values. These often are useful for discrete predictors. For continuous predictors, if the focus is on marginal effects, they often are not interesting. However, the at
argument can be used with continuous predictors to generate interesting averaged predicted values. For example, this would get predicted values integrating out random effects for a range of ages averaging (marginalizing) all other predictors / covariates.
ames < brmsmargins(
object = model,
at = data.frame(age = c(20, 30, 40, 50, 60)),
effects = "integrateoutRE")
ames$Summary
Installation
The package is not yet on CRAN, so to install, you must use the development version. To install, run:
remotes::install_github("JWiley/brmsmargins")
Learn More
There are three vignettes that introduce how to use the package for several scenarios.
 Fixed effects only models (also called single level models). This also is the best place to start learning about how to use the package. It includes a brief amount of motivation for why we would want to calculate marginal effects at all.
 Mixed effects models (also called multilevel models). This shows how to calculate marginal effects for mixed effects / multilevel models. There are runnable examples, but not much background.

Location scale models. Location scale models are models where both the location (e.g., mean) and scale (e.g., variance / residual standard deviation) are explicitly modeled as outcomes. These require use of distributional parameters
dpar
inbrms
. This vignette shows how to calculate marginal effects from location scale models for the scale part.