Compute marginal effects from statistical models and returns the result as tidy data frames. These data frames are ready to use with the 'ggplot2'-package. Marginal effects can be calculated for many different models. Interaction terms, splines and polynomial terms are also supported. The main functions are ggpredict(), ggemmeans() and ggeffect(). There is a generic plot()-method to plot the results using 'ggplot2'.
Results of regression models are typically presented as tables that are easy to understand. For more complex models that include interaction or quadratic / spline terms, tables with numbers are less helpful and difficult to interpret. In such cases, marginal effects are far easier to understand. In particular, the visualization of marginal effects allows to intuitively get the idea of how predictors and outcome are associated, even for complex models.
ggeffects computes marginal effects at the mean or average marginal effects from statistical models and returns the result as tidy data frames. These data frames are ready to use with the ggplot2-package.
Please visit https://strengejacke.github.io/ggeffects/ for documentation and vignettes. In case you want to file an issue or contribute in another way to the package, please follow this guide. For questions about the functionality, you may either contact me via email or also file an issue.
Marginal effects can be calculated for many different models. Currently supported model-objects are: betareg
, bglmer
, blmer
, brglm
, brmsfit
, clm
, clm2
, clmm
, coxph
, gam
(package mgcv), Gam
(package gam), gamlss
, gamm
, gamm4
, gee
, geeglm
, glm
, glm.nb
, glmer
, glmer.nb
, glmmTMB
, glmmPQL
, glmrob
, glmRob
, gls
, hurdle
, ivreg
, lm
, lm_robust
, lme
, lmer
, lmrob
, lmRob
, logistf
, lrm
, MixMod
, MCMCglmm
, multinom
, nlmer
, ols
, plm
, polr
, rlm
, rlmer
, rq
, rqss
, stanreg
, survreg
, svyglm
, svyglm.nb
, tobit
, truncreg
, vgam
, zeroinfl
and zerotrunc
. Other models not listed here are passed to a generic predict-function and might work as well, or maybe with ggeffect()
or ggemmeans()
, which effectively do the same as ggpredict()
.
Interaction terms, splines and polynomial terms are also supported. The main functions are ggpredict()
, ggemmeans()
and ggeffect()
. There is a generic plot()
-method to plot the results using ggplot2.
The returned data frames always have the same, consistent structure and column names, so it's easy to create ggplot-plots without the need to re-write the function call. x
and predicted
are the values for the x- and y-axis. conf.low
and conf.high
could be used as ymin
and ymax
aesthetics for ribbons to add confidence bands to the plot. group
can be used as grouping-aesthetics, or for faceting.
ggpredict()
requires at least one, but not more than three terms specified in the terms
-argument. Predicted values of the response, along the values of the first term are calucalted, optionally grouped by the other terms specified in terms
.
data(efc)
fit <- lm(barthtot ~ c12hour + neg_c_7 + c161sex + c172code, data = efc)
ggpredict(fit, terms = "c12hour")
#>
#> # Predicted values of Total score BARTHEL INDEX
#> # x = average number of hours of care per week
#>
#> x predicted std.error conf.low conf.high
#> 0 75.444 1.116 73.257 77.630
#> 15 71.644 0.965 69.753 73.535
#> 35 66.578 0.851 64.911 68.245
#> 50 62.779 0.852 61.108 64.449
#> 70 57.713 0.970 55.811 59.614
#> 85 53.913 1.122 51.713 56.113
#> 100 50.113 1.309 47.547 52.680
#> 120 45.047 1.591 41.929 48.166
#> 135 41.248 1.817 37.686 44.810
#> 170 32.382 2.373 27.732 37.033
#>
#> Adjusted for:
#> * neg_c_7 = 11.84
#> * c161sex = 1.76
#> * c172code = 1.97
A possible call to ggplot could look like this:
library(ggplot2)
mydf <- ggpredict(fit, terms = "c12hour")
ggplot(mydf, aes(x, predicted)) +
geom_line() +
geom_ribbon(aes(ymin = conf.low, ymax = conf.high), alpha = .1)
However, there is also a plot()
-method. This method uses convenient defaults, to easily create the most suitable plot for the marginal effects.
mydf <- ggpredict(fit, terms = "c12hour")
plot(mydf)
ggeffects has a plot()
-method with some convenient defaults, which allows quickly creating ggplot-objects.
With three variables, predictions can be grouped and faceted.
ggpredict(fit, terms = c("c12hour", "c172code", "c161sex"))
#>
#> # Predicted values of Total score BARTHEL INDEX
#> # x = average number of hours of care per week
#>
#> # low level of education
#> # [1] Male
#> x predicted std.error conf.low conf.high
#> 0 73.954 2.347 69.354 78.554
#> 45 62.556 2.208 58.228 66.883
#> 85 52.424 2.310 47.896 56.951
#> 170 30.893 3.085 24.847 36.939
#>
#> # low level of education
#> # [2] Female
#> x predicted std.error conf.low conf.high
#> 0 74.996 1.831 71.406 78.585
#> 45 63.597 1.603 60.456 66.738
#> 85 53.465 1.702 50.130 56.800
#> 170 31.934 2.606 26.827 37.042
#>
#> # intermediate level of education
#> # [1] Male
#> x predicted std.error conf.low conf.high
#> 0 74.673 1.845 71.055 78.290
#> 45 63.274 1.730 59.883 66.665
#> 85 53.142 1.911 49.397 56.887
#> 170 31.611 2.872 25.982 37.241
#>
#> # intermediate level of education
#> # [2] Female
#> x predicted std.error conf.low conf.high
#> 0 75.714 1.225 73.313 78.115
#> 45 64.315 0.968 62.418 66.213
#> 85 54.183 1.209 51.815 56.552
#> 170 32.653 2.403 27.943 37.362
#>
#> # high level of education
#> # [1] Male
#> x predicted std.error conf.low conf.high
#> 0 75.391 2.220 71.040 79.741
#> 45 63.992 2.176 59.727 68.258
#> 85 53.860 2.364 49.226 58.494
#> 170 32.330 3.257 25.946 38.713
#>
#> # high level of education
#> # [2] Female
#> x predicted std.error conf.low conf.high
#> 0 76.432 1.809 72.887 79.977
#> 45 65.034 1.712 61.679 68.388
#> 85 54.902 1.910 51.158 58.646
#> 170 33.371 2.895 27.697 39.045
#>
#> Adjusted for:
#> * neg_c_7 = 11.84
mydf <- ggpredict(fit, terms = c("c12hour", "c172code", "c161sex"))
ggplot(mydf, aes(x = x, y = predicted, colour = group)) +
stat_smooth(method = "lm", se = FALSE) +
facet_wrap(~facet)
plot()
works for this case, as well:
plot(mydf)
There are some more features, which are explained in more detail in the package-vignette.
Please follow this guide if you like to contribute to this package.
To install the latest development snapshot (see latest changes below), type following commands into the R console:
library(devtools)devtools::install_github("strengejacke/ggeffects")
Please note the package dependencies when installing from GitHub. The GitHub version of this package may depend on latest GitHub versions of my other packages, so you may need to install those first, if you encounter any problems. Here's the order for installing packages from GitHub:
sjlabelled → sjmisc → sjstats → ggeffects → sjPlot
To install the latest stable release from CRAN, type following command into the R console:
install.packages("ggeffects")
In case you want / have to cite my package, please use citation('ggeffects')
for citation information.
gamlss
, geeglm
(package geepack), lmrob
and glmrob
(package robustbase), ols
(package rms), rlmer
(package robustlmm), rq
and rqss
(package quantreg), tobit
(package AER), survreg
(package survival)terms = "predictor [1:10]"
) can now be changed with by
, e.g. terms = "predictor [1:10 by=.5]"
(see also vignette Marginal Effects at Specific Values).vcov.fun
in ggpredict()
) now also works for following model-objects: coxph
, plm
, polr
(and probably also lme
and gls
, not tested yet).ggpredict()
gets an interval
-argument, to compute prediction intervals instead of confidence intervals.plot.ggeffects()
now allows different horizontal and vertical jittering for rawdata
when jitter
is a numeric vector of length two.AsIs
-conversion from division of two variables as dependent variable, e.g. I(amount/frequency)
, now should work.ggpredict()
failed for MixMod
-objects when ci.lvl=NA
.ggemmeans()
now supports type = "fe.zi"
for glmmTMB-models, i.e. predicted values are conditioned on the fixed effects and the zero-inflation components of glmmTMB-models.ggpredict()
now supports MCMCglmm, ivreg and MixMod (package GLMMadaptive) models.ggemmeans()
now supports MCMCglmm and MixMod (package GLMMadaptive) models.ggpredict()
now computes confidence intervals for gam models (package gam).new_data()
, to create a data frame from all combinations of predictor values. This data frame typically can be used for the newdata
-argument in predict()
, in case it is necessary to quickly create an own data frame for this argument.ggpredict()
no longer stops when predicted values with confidence intervals for glmmTMB- and other zero-inflated models can't be computed with type = "fe.zi"
, and only returns the predicted values without confidence intervals.ggpredict()
fails to compute confidence intervals, a more informative error message is given.plot()
gets a connect.lines
-argument, to connect dots from plots with discrete x-axis.ggpredict()
did not work with glmmTMB- and other zero-inflated models, when type = "fe.zi"
and model- or zero-inflation formula had a polynomial term that was held constant (i.e. not part of the terms
-argument).type = "fe.zi"
could not be computed when the model contained polynomial terms and a very long formula (issue with deparse()
, cutting off very long formulas).plot()
-method put different spacing between groups when a numeric factor was used along the x-axis, where the factor levels where non equal-spaced.lm
in ggeffects()
.type = "fe"
and type = "re"
return population-level predictions for mixed effects models (lme4, glmmTMB). The difference is that type = "re"
also takes the random effect variances for prediction intervals into account. Predicted values at specific levels of random effect terms is described in the package-vignettes Marginal Effects for Random Effects Models and Marginal Effects at Specific Values.terms
-argument.plot()
. Use show_pals()
to show all available palettes.ggpredict()
and ggeffect()
now support brms-models with additional response information (like trial()
).ggpredict()
now supports Gam, glmmPQL, clmm, and zerotrunc-models.ggemmeans()
-function. Since this function is quite new, there still might be some bugs, though.ggemmeans()
to compute marginal effects by calling emmeans::emmeans()
.theme_ggeffects()
, which can be used with ggplot2::theme_set()
to set the ggeffects-theme as default plotting theme. This makes it easier to add further theme-modifications like sjPlot::legend_style()
or sjPlot::font_size()
.type = "sim"
) to ggpredict()
, currently for models of class glmmTMB and merMod.x.cat
is a new alias for the argument x.as.factor
.plot()
-method gets a ci.style
-argument, to define different styles for the confidence bands for numeric x-axis-terms.print()
-method gets a x.lab
-argument to print value labels instead of numeric values if x
is categorical.emm()
now also supports all prediction-types, like ggpredict()
.ggeffect()
, which did not work if data had variables with more that 8 digits (fractional part longer than 8 numbers).ppd = TRUE
.type = "fe.zi"
, which could mess up the correct order of predicted values for x
.type = "fe.zi"
or type = "re.zi"
, when first terms had the [all]
-tag.print()
-method for mixed effects models, when predictions were conditioned on all model terms and adjustment was only done for random effects (output-line "adjusted for").terms
included a factor and contrasts
were set to other values than contr.treatment
.glm
-object and heteroskedasticity-consistent covariance matrix estimation.condition
-argument was not always considered for some model types when calculating confidence intervals for predicted values.mo()
).[all]
tag, i.e. terms="... [all]"
).[all]
tag, i.e. terms="... [all]"
, will produce smoother plots.print()
-method, that - for larger data frames - only prints representative data rows. Use the n
-argument inside the print()
-method to force a specific number of rows to be printed.n
-tag for the terms
-argument in ggpredict()
and ggeffect()
, to give more flexibility according to how many values are used for "prettifying" large value ranges.sample
-tag for the terms
-argument in ggpredict()
and ggeffect()
, to pick a random sample of values for plotting.ggpredict()
and ggeffect()
now also return the standard error of predictions, if available.jitter
-argument in plot()
now also changes the amount of noise for plots of models with binary outcome (when rawdata = TRUE
).type="re"
and type="re.zi"
in general, and also for models with ar1
random effects structure.print()
-method, with a nicer print of the returned data frame. This method replaces the summary()
-method, which was removed.ggeffect()
now supports clm2
-models from the ordinal-package.ggpredict()
has improved support for coxph
-models from the survival-package (survival probabilities, cumulative hazards).type
-argument in ggpredict()
now has additional options, type = "fe.zi"
and type = "re.zi"
, to explicitely condition zero-inflated (mixed) models on their zero-inflation component.type
-argument in ggpredict()
now has additional options, type = "surv"
and type = "cumhaz"
, to plot probabilities of survival or cumulative hazards from coxph
-models.ggpredict()
gets arguments vcov.fun
, vcov.type
and vcov.args
to calculate robust standard errors for confidence intervals of predicted values. These are based on the various sandwich::vcov*()
-functions, hence robust standard errors can be calculated for all models that are supported by sandwich::vcov*()
.plot()
-method gets two arguments line.size
and dot.size
, to determine the size of the geoms.ci
-argument for the plot()
-method now also accepts the character values "dash"
and "dot"
to plot dashed or dotted lines as confidence bands.terms
-argument in ggpredict()
and ggeffect()
may also be a formula, which is more convenient for typing, but less flexible than specifying the terms as character vector with specific options.ggpredict()
now automatically back-transforms predictions to the response scale for model with log-transformed response.ggeffect()
and ggpredict()
now automatically set numeric vectors with 10 or more unique values to representative values (see rprs_values()
), if these are used as second or third value in the terms
-argument (to represent a grouping structure).ggeffect()
.rprs_values()
is now exported.pretty
-argument is deprecated, because prettifying values almost always makes sense - so this is done automatically.ggpredict()
now supports brmsfit
-objects with categorical-family.ggalleffect()
has been removed. ggeffect()
now plots effects for all model terms if terms = NULL
.gginteraction()
and ggpoly()
have been removed, as ggpredict()
and ggeffect()
are more efficient and generic for plotting interaction or polynomial terms.polr
, clm
, multinom
) for ggeffect()
.cbind()
in model formula.