Skip to contents

Adjusted predictions from regression models

Source: R/data_frame_methods.R, R/ggaverage.R, R/ggeffect.R, and 2 more
ggpredict.Rd

After fitting a model, it is useful generate model-based estimates (expected values, or adjusted predictions) of the response variable for different combinations of predictor values. Such estimates can be used to make inferences about relationships between variables.

The ggeffects package computes marginal means and adjusted predicted values for the response, at the margin of specific values or levels from certain model terms. The package is built around three core functions: predict_response() (understanding results), test_predictions() (testing results for statistically significant differences) and plot() (communicate results).

By default, adjusted predictions or marginal means are by returned on the response scale, which is the easiest and most intuitive scale to interpret the results. There are other options for specific models as well, e.g. with zero-inflation component (see documentation of the type-argument). The result is returned as consistent data frame, which is nicely printed by default. plot() can be used to easily create figures.

The main function to calculate marginal means and adjusted predictions is predict_response(). In previous versions of ggeffects, the functions ggpredict(), ggemmeans(), ggeffect() and ggaverage() were used to calculate marginal means and adjusted predictions. These functions are still available, but predict_response() as a "wrapper" around these functions is the preferred way to do this now.

Usage

# S3 method for class 'ggeffects'
as.data.frame(
  x,
  row.names = NULL,
  optional = FALSE,
  ...,
  stringsAsFactors = FALSE,
  terms_to_colnames = FALSE
)

ggaverage(
  model,
  terms,
  ci_level = 0.95,
  type = "fixed",
  typical = "mean",
  condition = NULL,
  back_transform = TRUE,
  vcov_fun = NULL,
  vcov_type = NULL,
  vcov_args = NULL,
  weights = NULL,
  verbose = TRUE,
  ...
)

ggeffect(model, terms, ci_level = 0.95, verbose = TRUE, ci.lvl = ci_level, ...)

ggemmeans(
  model,
  terms,
  ci_level = 0.95,
  type = "fixed",
  typical = "mean",
  condition = NULL,
  back_transform = TRUE,
  vcov_fun = NULL,
  vcov_type = NULL,
  vcov_args = NULL,
  interval = "confidence",
  verbose = TRUE,
  ci.lvl = ci_level,
  back.transform = back_transform,
  ...
)

ggpredict(
  model,
  terms,
  ci_level = 0.95,
  type = "fixed",
  typical = "mean",
  condition = NULL,
  back_transform = TRUE,
  vcov_fun = NULL,
  vcov_type = NULL,
  vcov_args = NULL,
  interval,
  verbose = TRUE,
  ci.lvl = ci_level,
  back.transform = back_transform,
  vcov.fun = vcov_fun,
  vcov.type = vcov_type,
  vcov.args = vcov_args,
  ...
)

Arguments

x

An object of class ggeffects, as returned by predict_response(), ggpredict(), ggeffect(), ggaverage() or ggemmeans().

row.names

NULL or a character vector giving the row names for the data frame. Missing values are not allowed.

optional

logical. If TRUE, setting row names and converting column names (to syntactic names: see make.names) is optional. Note that all of R's base package as.data.frame() methods use optional only for column names treatment, basically with the meaning of data.frame(*, check.names = !optional). See also the make.names argument of the matrix method.

...

Arguments are passed down to ggpredict() (further down to predict()) or ggemmeans() (and thereby to emmeans::emmeans()), If type = "simulate", ... may also be used to set the number of simulation, e.g. nsim = 500. When calling ggeffect(), further arguments passed down to effects::Effect().

stringsAsFactors

logical: should the character vector be converted to a factor?

terms_to_colnames

Logical, if TRUE, standardized column names (like "x", "group" or "facet") are replaced by the variable names of the focal predictors specified in terms.

model

A model object, or a list of model objects.

terms

Names of those terms from model, for which predictions should be displayed (so called focal terms). Can be:

  • A character vector, specifying the names of the focal terms. This is the preferred and probably most flexible way to specify focal terms, e.g. terms = "x [40:60]", to calculate predictions for the values 40 to 60.

  • A list, where each element is a named vector, specifying the focal terms and their values. This is the "classical" R way to specify focal terms, e.g. list(x = 40:60).

  • A formula, e.g. terms = ~ x + z, which is internally converted to a character vector. This is probably the least flexible way, as you cannot specify representative values for the focal terms.

  • A data frame representing a "data grid" or "reference grid". Predictions are then made for all combinations of the variables in the data frame.

terms at least requires one variable name. The maximum length is four terms, where the second to fourth term indicate the groups, i.e. predictions of the first term are grouped at meaningful values or levels of the remaining terms (see values_at()). It is also possible to define specific values for focal terms, at which adjusted predictions should be calculated (see details below). All remaining covariates that are not specified in terms are "marginalized", see the margin argument in ?predict_response. See also argument condition to fix non-focal terms to specific values, and argument typical for ggpredict() or ggemmeans().

ci_level

Numeric, the level of the confidence intervals. Use ci_level = NA if confidence intervals should not be calculated (for instance, due to computation time). Typically, confidence intervals are based on the returned standard errors for the predictions, assuming a t- or normal distribution (based on the model and the available degrees of freedom, i.e. roughly +/- 1.96 * SE). See introduction of this vignette for more details.

type

Character, indicating whether predictions should be conditioned on specific model components or not. Consequently, most options only apply for survival models, mixed effects models and/or models with zero-inflation (and their Bayesian counter-parts); only exception is type = "simulate", which is available for some other model classes as well (which respond to simulate()).

Note 1: For brmsfit-models with zero-inflation component, there is no type = "zero_inflated" nor type = "zi_random"; predicted values for these models always condition on the zero-inflation part of the model. The same is true for MixMod-models from GLMMadaptive with zero-inflation component (see 'Details').

Note 2: If margin = "empirical", or when calling ggaverage() respectively, (i.e. counterfactual predictions), the type argument is handled differently. It is set to "response" by default, but usually accepts all possible options from the type-argument of the model's respective predict() method. E.g., passing a glm object would allow the options "response", "link", and "terms". For models with zero-inflation component, the below mentioned options "fixed", "zero_inflated" and "zi_prob" can also be used and will be "translated" into the corresponding type option of the model's respective predict()-method.

  • "fixed" (or "count")

    Predicted values are conditioned on the fixed effects or conditional model only (for mixed models: predicted values are on the population-level and confidence intervals are returned, i.e. re.form = NA when calling predict()). For instance, for models fitted with zeroinfl from pscl, this would return the predicted mean from the count component (without zero-inflation). For models with zero-inflation component, this type calls predict(..., type = "link") (however, predicted values are back-transformed to the response scale, i.e. the conditional mean of the response).

  • "random"

    This only applies to mixed models, and type = "random" does not condition on the zero-inflation component of the model. type = "random" still returns population-level predictions, however, conditioned on random effects and considering individual level predictions, i.e. re.form = NULL when calling predict(). This may affect the returned predicted values, depending on whether REML = TRUE or REML = FALSE was used for model fitting. Furthermore, unlike type = "fixed", intervals also consider the uncertainty in the variance parameters (the mean random effect variance, see Johnson et al. 2014 for details) and hence can be considered as prediction intervals. For models with zero-inflation component, this type calls predict(..., type = "link") (however, predicted values are back-transformed to the response scale).

    To get predicted values for each level of the random effects groups, add the name of the related random effect term to the terms-argument (for more details, see this vignette).

  • "zero_inflated" (or "zi")

    Predicted values are conditioned on the fixed effects and the zero-inflation component. For instance, for models fitted with zeroinfl from pscl, this would return the predicted (or expected) response (mu*(1-p)), and for glmmTMB, this would return the expected response mu*(1-p) without conditioning on random effects (i.e. random effect variances are not taken into account for the confidence intervals). For models with zero-inflation component, this type calls predict(..., type = "response"). See 'Details'.

  • "zi_random" (or "zero_inflated_random")

    Predicted values are conditioned on the zero-inflation component and take the random effects uncertainty into account. For models fitted with glmmTMB(), hurdle() or zeroinfl(), this would return the expected value mu*(1-p). For glmmTMB, prediction intervals also consider the uncertainty in the random effects variances. This type calls predict(..., type = "response"). See 'Details'.

  • "zi_prob"

    Predicted zero-inflation probability. For glmmTMB models with zero-inflation component, this type calls predict(..., type = "zlink"); models from pscl call predict(..., type = "zero") and for GLMMadaptive, predict(..., type = "zero_part") is called.

  • "simulate"

    Predicted values and confidence resp. prediction intervals are based on simulations, i.e. calls to simulate(). This type of prediction takes all model uncertainty into account, including random effects variances. Currently supported models are objects of class lm, glm, glmmTMB, wbm, MixMod and merMod. See ... for details on number of simulations.

  • "survival" and "cumulative_hazard"

    Applies only to coxph-objects from the survial-package and calculates the survival probability or the cumulative hazard of an event.

When margin = "empirical" (or when calling ggaverage()), the type argument accepts all values from the type-argument of the model's respective predict()-method.

typical

Character vector, naming the function to be applied to the covariates (non-focal terms) over which the effect is "averaged". The default is "mean". Can be "mean", "weighted.mean", "median", "mode" or "zero", which call the corresponding R functions (except "mode", which calls an internal function to compute the most common value); "zero" simply returns 0. By default, if the covariate is a factor, only "mode" is applicable; for all other values (including the default, "mean") the reference level is returned. For character vectors, only the mode is returned. You can use a named vector to apply different functions to integer, numeric and categorical covariates, e.g. typical = c(numeric = "median", factor = "mode"). If typical is "weighted.mean", weights from the model are used. If no weights are available, the function falls back to "mean". Note that this argument is ignored for predict_response(), because the margin argument takes care of this.

condition

Named character vector, which indicates covariates that should be held constant at specific values. Unlike typical, which applies a function to the covariates to determine the value that is used to hold these covariates constant, condition can be used to define exact values, for instance condition = c(covariate1 = 20, covariate2 = 5). See 'Examples'.

back_transform

Logical, if TRUE (the default), predicted values for log-, log-log, exp, sqrt and similar transformed responses will be back-transformed to original response-scale. See insight::find_transformation() for more details.

vcov_fun

Variance-covariance matrix used to compute uncertainty estimates (e.g., for confidence intervals based on robust standard errors). This argument accepts a covariance matrix, a function which returns a covariance matrix, or a string which identifies the function to be used to compute the covariance matrix.

  • A (variance-covariance) matrix

  • A function which returns a covariance matrix (e.g., stats::vcov())

  • A string which indicates the estimation type for the heteroscedasticity-consistent variance-covariance matrix, e.g. vcov_fun = "HC0". Possible values are "HC0", "HC1", "HC2", "HC3", "HC4", "HC4m", and "HC5", which will then call the vcovHC()-function from the sandwich package, using the specified type. Further possible values are "CR0", "CR1", "CR1p", "CR1S", "CR2", and "CR3", which will call the vcovCR()-function from the clubSandwich package.

  • A string which indicates the name of the vcov*()-function from the sandwich or clubSandwich packages, e.g. vcov_fun = "vcovCL", which is used to compute (cluster) robust standard errors for predictions.

If NULL, standard errors (and confidence intervals) for predictions are based on the standard errors as returned by the predict()-function. Note that probably not all model objects that work with predict_response() are also supported by the sandwich or clubSandwich packages.

See details in this vignette.

vcov_type

Character vector, specifying the estimation type for the robust covariance matrix estimation (see ?sandwich::vcovHC or ?clubSandwich::vcovCR for details). Only used when vcov_fun is a character string indicating one of the functions from those packages. When vcov_fun is a function, a possible type argument must be provided via the vcov_args argument.

vcov_args

List of named vectors, used as additional arguments that are passed down to vcov_fun.

weights

This argument is used in two different ways, depending on the margin argument.

  • When margin = "empirical", weights can either be a character vector, naming the weigthing variable in the data, or a vector of weights (of same length as the number of observations in the data). This variable will be used to weight adjusted predictions.

  • When margin = "marginalmeans", weights must be a character vector and is passed to emmeans::emmeans(), specifying weights to use in averaging non-focal categorical predictors. See https://rvlenth.github.io/emmeans/reference/emmeans.html for details.

verbose

Toggle messages or warnings.

ci.lvl, vcov.fun, vcov.type, vcov.args, back.transform

Deprecated arguments. Please use ci_level, vcov_fun, vcov_type, vcov_args and back_transform instead.

interval

Type of interval calculation, can either be "confidence" (default) or "prediction". May be abbreviated. Unlike confidence intervals, prediction intervals include the residual variance (sigma^2) to account for the uncertainty of predicted values. For mixed models, interval = "prediction" is the default for type = "random". When type = "fixed", the default is interval = "confidence". Note that prediction intervals are not available for all models, but only for models that work with insight::get_sigma(). For Bayesian models, when interval = "confidence", predictions are based on posterior draws of the linear predictor rstantools::posterior_epred(). If interval = "prediction", rstantools::posterior_predict() is called.

Value

A data frame (with ggeffects class attribute) with consistent data columns:

  • "x": the values of the first term in terms, used as x-position in plots.

  • "predicted": the predicted values of the response, used as y-position in plots.

  • "std.error": the standard error of the predictions. Note that the standard errors are always on the link-scale, and not back-transformed for non-Gaussian models!

  • "conf.low": the lower bound of the confidence interval for the predicted values.

  • "conf.high": the upper bound of the confidence interval for the predicted values.

  • "group": the grouping level from the second term in terms, used as grouping-aesthetics in plots.

  • "facet": the grouping level from the third term in terms, used to indicate facets in plots.

    The estimated marginal means (or predicted values) are always on the response scale!

    For proportional odds logistic regression (see ?MASS::polr) resp. cumulative link models (e.g., see ?ordinal::clm), an additional column "response.level" is returned, which indicates the grouping of predictions based on the level of the model's response.

    Note that for convenience reasons, the columns for the intervals are always named "conf.low" and "conf.high", even though for Bayesian models credible or highest posterior density intervals are returned.

    There is an as.data.frame() method for objects of class ggeffects, which has an terms_to_colnames argument, to use the term names as column names instead of the standardized names "x" etc.

Details

Please see ?predict_response for details and examples.