This vignette shows how to add support for new models and add new functionality for supported models.

## Support a new model type

It is very easy to add support for new models in marginaleffects. All we need is to set a global option and define 4 very simple functions.

If you add support for a class of models produced by a CRAN package, please consider submitting your code for inclusion in the package: https://github.com/vincentarelbundock/marginaleffects

If you add support for a class of models produced by a package hosted elsewhere than CRAN, you can submit it for inclusion in the unsupported user-submitted library of extensions: Currently

The rest of this section illustrates how to add support for a very simple lm_manual model.

### Fit function

To begin, we define a function which fits a model. Normally, this function will be supplied by a modeling package published on CRAN. Here, we create a function called lm_manual(), which estimates a linear regression model using simple linear algebra operates:

lm_manual <- function(f, data, ...) {
# design matrix
X <- model.matrix(f, data = data)
# response matrix
Y <- data[[as.character(f)]]
# coefficients
b <- solve(crossprod(X)) %*% crossprod(X, Y)
Yhat <- X %*% b
# variance-covariance matrix
e <- Y - Yhat
df <- nrow(X) - ncol(X)
s2 <- sum(e^2) / df
V <- s2 * solve(crossprod(X))
# model object
out <- list(
d = data,
f = f,
X = X,
Y = Y,
V = V,
b = b)
# class name: lm_manual
class(out) <- c("lm_manual", "list")
return(out)
}

Important: The custom fit function must assign a new class name to the object it returns. In the example above, the model is assigned to be of class lm_manual (see the penultimate line of code in the function).

Our new function replicates the results of lm():

model <- lm_manual(mpg ~ hp + drat, data = mtcars)
model$b #> [,1] #> (Intercept) 10.78986122 #> hp -0.05178665 #> drat 4.69815776 model_lm <- lm(mpg ~ hp + drat, data = mtcars) coef(model_lm) #> (Intercept) hp drat #> 10.78986122 -0.05178665 4.69815776 ### marginaleffects extension To extend support in marginaleffects, the first step is to tell the package that our new class is supported. We do this by defining a global option: library(marginaleffects) options("marginaleffects_model_classes" = "lm_manual") Then, we define 4 methods: 1. get_coef() • Mandatory arguments: model, ... • Returns: named vector of parameters (coefficients). 2. set_coef() • Mandatory arguments: model, coefs (named vector of coefficients), ... • Returns: A new model object in which the original coefficients were replaced by the new vector. • Example 3. get_vcov() • Mandatory arguments: model, .... • Optional arguments: vcov • Returns: A named square variance-covariance matrix. 4. get_predict() • Mandatory arguments: model, newdata (data frame), ... • Option arguments: type and other model-specific arguments. • Returns: A data frame with two columns: a unique rowid and a column of estimate values. Note that each of these methods will be named with the suffix .lm_manual to indicate that they should be used whenever marginaleffects needs to process an object of class lm_manual. get_coef.lm_manual <- function(model, ...) { b <- model$b
b <- setNames(as.vector(b), row.names(b))
return(b)
}

set_coef.lm_manual <- function(model, coefs, ...) {
out <- model
out$b <- coefs return(out) } get_vcov.lm_manual <- function(model, ...) { return(model$V)
}

get_predict.lm_manual <- function(model, newdata, ...) {
newX <- model.matrix(model$f, data = newdata) Yhat <- newX %*% model$b
out <- data.frame(
rowid = seq_len(nrow(Yhat)),
estimate = as.vector(Yhat))
return(out)
}

The methods we just defined work as expected:

get_coef(model)
#> (Intercept)          hp        drat
#> 10.78986122 -0.05178665  4.69815776

get_vcov(model)
#>             (Intercept)            hp         drat
#> (Intercept) 25.78356135 -3.054007e-02 -5.836030687
#> hp          -0.03054007  8.635615e-05  0.004969385
#> drat        -5.83603069  4.969385e-03  1.419990359

#>   rowid estimate
#> 1     1 23.41614
#> 2     2 23.41614
#> 3     3 24.06161
#> 4     4 19.56366
#> 5     5 16.52639
#> 6     6 18.31918

Now we can use the avg_slopes function:

avg_slopes(model, newdata = mtcars, variables = c("hp", "drat"))
#>
#>  Term Estimate Std. Error      z   Pr(>|z|)  2.5 %   97.5 %
#>    hp -0.05179   0.009293 -5.573 2.5050e-08 -0.070 -0.03357
#>  drat  4.69816   1.191633  3.943 8.0596e-05  2.363  7.03372
#>
#> Prediction type:  response
#> Columns: type, term, estimate, std.error, statistic, p.value, conf.low, conf.high

predictions(model, newdata = mtcars) |>
#>
#>  Estimate Std. Error     z   Pr(>|z|) 2.5 % 97.5 %
#>     23.42     0.6711 34.89 < 2.22e-16 22.10  24.73
#>     23.42     0.6711 34.89 < 2.22e-16 22.10  24.73
#>     24.06     0.7204 33.40 < 2.22e-16 22.65  25.47
#>     19.56     0.9988 19.59 < 2.22e-16 17.61  21.52
#>     16.53     0.7354 22.47 < 2.22e-16 15.09  17.97
#>     18.32     1.3433 13.64 < 2.22e-16 15.69  20.95
#>
#> Prediction type:  response
#> Columns: rowid, type, estimate, std.error, statistic, p.value, conf.low, conf.high, mpg, cyl, disp, hp, drat, wt, qsec, vs, am, gear, carb

Note that, for custom model, we typically have to supply values for the newdata and variables arguments explicitly.

## Modify or extend supported models

Let’s say you want to estimate a model using the mclogit::mblogit function. That package is already supported by marginaleffects, but you want to use a type (scale) of predictions that is not currently supported: a “centered link scale.”

To achieve this, we would need to override the get_predict.mblogit() method. However, it can be unsafe to reassign methods supplied by a package that we loaded with library. To be safe, we assign a new model class to our object (“customclass”) which will inherit from mblogit. Then, we define a get_predict.customclass method to make our new kinds of preditions.

library(mclogit)
library(data.table)

model <- mblogit(
factor(gear) ~ am + mpg,
data = mtcars,
trace = FALSE)

Tell marginaleffects that we are adding support for a new class model models, and assign a new inherited class name to a duplicate of the model object:

options("marginaleffects_model_classes" = "customclass")

model_custom <- model

class(model_custom) <- c("customclass", class(model))

Define a new get_predict.customclass method. We use the default predict() function to obtain predictions. Since this is a multinomial model, predict() returns a matrix of predictions with one column per level of the response variable.

Our new get_predict.customclass method takes this matrix of predictions, modifies it, and reshapes it to return a data frame with three columns: rowid, group, and estimate:

get_predict.customclass <- function(model, newdata, ...) {
out <- predict(model, newdata = newdata, type = "link")
out <- cbind(0, out)
colnames(out) <- dimnames(model$D)[][] out <- out - rowMeans(out) out <- as.data.frame(out) out$rowid <- seq_len(nrow(out))
out <- data.table(out)
out <- melt(
out,
id.vars = "rowid",
value.name = "estimate",
variable.name = "group")
}

Finally, we can call any slopes function and obtain results. Notice that our object of class customclass now produces different results than the default mblogit object:

avg_predictions(model)
#>
#>  Group Estimate Std. Error      z   Pr(>|z|)   2.5 % 97.5 %
#>      3   0.4688    0.04440 10.558 < 2.22e-16 0.38173 0.5558
#>      4   0.3750    0.06764  5.544 2.9505e-08 0.24244 0.5076
#>      5   0.1563    0.05103  3.062  0.0021976 0.05624 0.2563
#>
#> Prediction type:  response
#> Columns: type, group, estimate, std.error, statistic, p.value, conf.low, conf.high

avg_predictions(model_custom)
#>
#>  Group Estimate Std. Error          z Pr(>|z|) 2.5 % 97.5 %
#>      3   -1.418       2525 -0.0005614  0.99955 -4950   4947
#>      4    6.365       1779  0.0035776  0.99715 -3480   3493
#>      5   -4.947       3074 -0.0016094  0.99872 -6030   6020
#>
#> Prediction type:  response
#> Columns: type, group, estimate, std.error, statistic, p.value, conf.low, conf.high