Plotting Marginal Effects of Regression Models

2019-09-29

This document describes how to plot marginal effects of various regression models, using the plot_model() function. plot_model() is a generic plot-function, which accepts many model-objects, like lm, glm, lme, lmerMod etc.

plot_model() allows to create various plot tyes, which can be defined via the type-argument. The default is type = "fe", which means that fixed effects (model coefficients) are plotted. To plot marginal effects, call plot_model() with:

• type = "pred" to plot predicted values (marginal effects) for specific model terms.
• type = "eff", which is similar to type = "pred", however, discrete predictors are held constant at their proportions (not reference level).
• type = "int" to plot marginal effects of interaction terms.

To plot marginal effects of regression models, at least one model term needs to be specified for which the effects are computed. It is also possible to compute marginal effects for model terms, grouped by the levels of another model’s predictor. The function also allows plotting marginal effects for two- or three-way-interactions, however, this is shown in a different vignette.

plot_model() supports labelled data and automatically uses variable and value labels to annotate the plot. This works with most regression modelling functions.

Marginal effects

plot_model(type = "pred") computes predicted values for all possible levels and values from a model’s predictors. In the simplest case, a fitted model is passed as first argument, followed by the type argument and the term in question as terms argument:

library(sjPlot)
library(ggplot2)
data(efc)
theme_set(theme_sjplot())

fit <- lm(barthtot ~ c12hour + neg_c_7 + c161sex + c172code, data = efc)

plot_model(fit, type = "pred", terms = "c12hour") The plot shows the predicted values for the response at each value from the term c12hour.

Marginal effects for different groups

The terms-argument accepts up to three model terms, where the second and third term indicate grouping levels. This allows predictions for the term in question at different levels for other model terms: A second grouping structure can be defined, which will create a plot with multiple panels in grid layout: Marginal effects at specific values or levels

The terms-argument not only defines the model terms of interest, but each model term that defines the grouping structure can be limited to certain values. This allows to compute and plot marginal effects for terms at specific values only. To define these values, put them in square brackets directly after the term name: terms = c("c12hour [30, 50, 80]", "c172code [1,3]") Note that in the above plot, although the values 30, 50 and 80 only are selected from c12hour, the continuous scale automatically adds panel grids every 5 units along the x-axis.

Defining own values is especially useful when variables are, for instance, log-transformed. plot_model() then typically only uses the range of the log-transformed variable, which is in most cases not what we want. In such situation, specify the range in the terms-argument.  The brackets in the terms-argument also accept the name of a valid function, to (back-)transform predicted valued. In this example, an alternative would be to specify that values should be exponentiated, which is indicated by [exp] in the terms-argument: Polynomial terms and splines

The function also works for models with polynomial terms or splines. Following code reproduces the plot from ?splines::bs: Different constant values for factors

Model predictions are based on all possible combinations of the model terms, which are - roughly speaking - created using expand.grid(). For the terms in question, all values are used for combinations. All other model predictors that are not specified in the terms-argument, are held constant (which is achieved with sjstats::typical_value()). By default, continuous variables are set to their mean, while factors are set to their reference level. However, one may want to set factors to their proportions instead of reference level. E.g., a factor gender with value 0 for female and value 1 for male persons, would be set to 0 when marginal effects are computed with type = "pred". But if 40% of the sample are female persons, another possibility to hold this factor constant is to use the value .4 (reflecting the proportion of 40%). If this is required, use type = "eff", which internally does not call predict() to compute marginal effects, but rather effects::effect(). Interaction terms

Plotting interaction terms are described in a separate vignette.