get_predicted {insight}

Model predictions (robust) and their confidence intervals

Description

The get_predicted() function is a robust, flexible and user-friendly alternative to base R predict() function. Additional features and advantages include availability of uncertainty intervals (CI), bootstrapping, a more intuitive API and the support of more models than base R's predict() function. However, although the interface are simplified, it is still very important to read the documentation of the arguments. This is because making "predictions" (a lose term for a variety of things) is a non-trivial process, with lots of caveats and complications. Read the 'Details' section for more information.

get_predicted_ci() returns the confidence (or prediction) interval (CI) associated with predictions made by a model. This function can be called separately on a vector of predicted values. get_predicted() usually returns confidence intervals (included as attribute, and accessible via the as.data.frame() method) by default. It is preferred to rely on the get_predicted() function for standard errors and confidence intervals - use get_predicted_ci() only if standard errors and confidence intervals are not available otherwise.

Usage

get_predicted(x, ...)

## Default S3 method:
get_predicted(
  x,
  data = NULL,
  predict = "expectation",
  ci = NULL,
  ci_type = "confidence",
  ci_method = NULL,
  dispersion_method = "sd",
  vcov = NULL,
  vcov_args = NULL,
  verbose = TRUE,
  ...
)

## S3 method for class 'lm'
get_predicted(
  x,
  data = NULL,
  predict = "expectation",
  ci = NULL,
  iterations = NULL,
  verbose = TRUE,
  ...
)

## S3 method for class 'stanreg'
get_predicted(
  x,
  data = NULL,
  predict = "expectation",
  iterations = NULL,
  ci = NULL,
  ci_method = NULL,
  include_random = "default",
  include_smooth = TRUE,
  verbose = TRUE,
  ...
)

## S3 method for class 'gam'
get_predicted(
  x,
  data = NULL,
  predict = "expectation",
  ci = NULL,
  include_random = TRUE,
  include_smooth = TRUE,
  iterations = NULL,
  verbose = TRUE,
  ...
)

## S3 method for class 'lmerMod'
get_predicted(
  x,
  data = NULL,
  predict = "expectation",
  ci = NULL,
  ci_method = NULL,
  include_random = "default",
  iterations = NULL,
  verbose = TRUE,
  ...
)

## S3 method for class 'principal'
get_predicted(x, data = NULL, ...)

Arguments

x

A statistical model (can also be a data.frame, in which case the second argument has to be a model).

...

Other argument to be passed, for instance to get_predicted_ci().

data

An optional data frame in which to look for variables with which to predict. If omitted, the data used to fit the model is used. Visualization matrices can be generated using get_datagrid().

predict

string or NULL

• "link" returns predictions on the model's link-scale (for logistic models, that means the log-odds scale) with a confidence interval (CI).

• "expectation" (default) also returns confidence intervals, but this time the output is on the response scale (for logistic models, that means probabilities).

• "prediction" also gives an output on the response scale, but this time associated with a prediction interval (PI), which is larger than a confidence interval (though it mostly make sense for linear models).

• "classification" only differs from "prediction" for binomial models where it additionally transforms the predictions into the original response's type (for instance, to a factor).

• Other strings are passed directly to the type argument of the predict() method supplied by the modelling package.

• When predict = NULL, alternative arguments such as type will be captured by the ... ellipsis and passed directly to the predict() method supplied by the modelling package. Note that this might result in conflicts with multiple matching type arguments - thus, the recommendation is to use the predict argument for those values.

• Notes: You can see the 4 options for predictions as on a gradient from "close to the model" to "close to the response data": "link", "expectation", "prediction", "classification". The predict argument modulates two things: the scale of the output and the type of certainty interval. Read more about in the Details section below.

ci

The interval level. Default is NULL, to be fast even for larger models. Set the interval level to an explicit value, e.g. 0.95, for ⁠95%⁠ CI).

ci_type

Can be "prediction" or "confidence". Prediction intervals show the range that likely contains the value of a new observation (in what range it would fall), whereas confidence intervals reflect the uncertainty around the estimated parameters (and gives the range of the link; for instance of the regression line in a linear regressions). Prediction intervals account for both the uncertainty in the model's parameters, plus the random variation of the individual values. Thus, prediction intervals are always wider than confidence intervals. Moreover, prediction intervals will not necessarily become narrower as the sample size increases (as they do not reflect only the quality of the fit). This applies mostly for "simple" linear models (like lm), as for other models (e.g., glm), prediction intervals are somewhat useless (for instance, for a binomial model for which the dependent variable is a vector of 1s and 0s, the prediction interval is... ⁠[0, 1]⁠).

ci_method

The method for computing p values and confidence intervals. Possible values depend on model type.

• NULL uses the default method, which varies based on the model type.

• Most frequentist models: "wald" (default), "residual" or "normal".

• Bayesian models: "quantile" (default), "hdi", "eti", and "spi".

• Mixed effects lme4 models: "wald" (default), "residual", "normal", "satterthwaite", and "kenward-roger".

See get_df() for details.

dispersion_method

Bootstrap dispersion and Bayesian posterior summary: "sd" or "mad".

vcov

Variance-covariance matrix used to compute uncertainty estimates (e.g., for 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 covariance matrix

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

• A string which indicates the kind of uncertainty estimates to return.

  • Heteroskedasticity-consistent: "vcovHC", "HC", "HC0", "HC1", "HC2", "HC3", "HC4", "HC4m", "HC5". See ?sandwich::vcovHC

  • Cluster-robust: "vcovCR", "CR0", "CR1", "CR1p", "CR1S", "CR2", "CR3". See ?clubSandwich::vcovCR()

  • Bootstrap: "vcovBS", "xy", "residual", "wild", "mammen", "webb". See ?sandwich::vcovBS

  • Other sandwich package functions: "vcovHAC", "vcovPC", "vcovCL", "vcovPL".

vcov_args

List of arguments to be passed to the function identified by the vcov argument. This function is typically supplied by the sandwich or clubSandwich packages. Please refer to their documentation (e.g., ?sandwich::vcovHAC) to see the list of available arguments. If no estimation type (argument type) is given, the default type for "HC" (or "vcovHC") equals the default from the sandwich package; for type "CR" (or "vcoCR"), the default is set to "CR3".

verbose

Toggle warnings.

iterations

For Bayesian models, this corresponds to the number of posterior draws. If NULL, will return all the draws (one for each iteration of the model). For frequentist models, if not NULL, will generate bootstrapped draws, from which bootstrapped CIs will be computed. Iterations can be accessed by running as.data.frame(..., keep_iterations = TRUE) on the output.

include_random

If "default", include all random effects in the prediction, unless random effect variables are not in the data. If TRUE, include all random effects in the prediction (in this case, it will be checked if actually all random effect variables are in data). If FALSE, don't take them into account. Can also be a formula to specify which random effects to condition on when predicting (passed to the re.form argument). If include_random = TRUE and data is provided, make sure to include the random effect variables in data as well.

include_smooth

For General Additive Models (GAMs). If FALSE, will fix the value of the smooth to its average, so that the predictions are not depending on it. (default), mean(), or bayestestR::map_estimate().

Details

In insight::get_predicted(), the predict argument jointly modulates two separate concepts, the scale and the uncertainty interval.

Value

The fitted values (i.e. predictions for the response). For Bayesian or bootstrapped models (when iterations != NULL), iterations (as columns and observations are rows) can be accessed via as.data.frame().

Confidence Interval (CI) vs. Prediction Interval (PI))

• Linear models - lm(): For linear models, prediction intervals (predict="prediction") show the range that likely contains the value of a new observation (in what range it is likely to fall), whereas confidence intervals (predict="expectation" or predict="link") reflect the uncertainty around the estimated parameters (and gives the range of uncertainty of the regression line). In general, Prediction Intervals (PIs) account for both the uncertainty in the model's parameters, plus the random variation of the individual values. Thus, prediction intervals are always wider than confidence intervals. Moreover, prediction intervals will not necessarily become narrower as the sample size increases (as they do not reflect only the quality of the fit, but also the variability within the data).

• Generalized Linear models - glm(): For binomial models, prediction intervals are somewhat useless (for instance, for a binomial (Bernoulli) model for which the dependent variable is a vector of 1s and 0s, the prediction interval is... ⁠[0, 1]⁠).

Link scale vs. Response scale

When users set the predict argument to "expectation", the predictions are returned on the response scale, which is arguably the most convenient way to understand and visualize relationships of interest. When users set the predict argument to "link", predictions are returned on the link scale, and no transformation is applied. For instance, for a logistic regression model, the response scale corresponds to the predicted probabilities, whereas the link-scale makes predictions of log-odds (probabilities on the logit scale). Note that when users select predict="classification" in binomial models, the get_predicted() function will first calculate predictions as if the user had selected predict="expectation". Then, it will round the responses in order to return the most likely outcome.

Heteroscedasticity consistent standard errors

The arguments vcov and vcov_args can be used to calculate robust standard errors for confidence intervals of predictions. These arguments, when provided in get_predicted(), are passed down to get_predicted_ci(), thus, see the related documentation there for more details.

Bayesian and Bootstrapped models and iterations

For predictions based on multiple iterations, for instance in the case of Bayesian models and bootstrapped predictions, the function used to compute the centrality (point-estimate predictions) can be modified via the centrality_function argument. For instance, get_predicted(model, centrality_function = stats::median). The default is mean. Individual draws can be accessed by running iter <- as.data.frame(get_predicted(model)), and their iterations can be reshaped into a long format by bayestestR::reshape_iterations(iter).

See Also

get_datagrid()

Examples

data(mtcars)
x <- lm(mpg ~ cyl + hp, data = mtcars)

predictions <- get_predicted(x, ci = 0.95)
predictions

# Options and methods ---------------------
get_predicted(x, predict = "prediction")

# Get CI
as.data.frame(predictions)

# Bootstrapped
as.data.frame(get_predicted(x, iterations = 4))
# Same as as.data.frame(..., keep_iterations = FALSE)
summary(get_predicted(x, iterations = 4))

# Different prediction types ------------------------
data(iris)
data <- droplevels(iris[1:100, ])

# Fit a logistic model
x <- glm(Species ~ Sepal.Length, data = data, family = "binomial")

# Expectation (default): response scale + CI
pred <- get_predicted(x, predict = "expectation", ci = 0.95)
head(as.data.frame(pred))

# Prediction: response scale + PI
pred <- get_predicted(x, predict = "prediction", ci = 0.95)
head(as.data.frame(pred))

# Link: link scale + CI
pred <- get_predicted(x, predict = "link", ci = 0.95)
head(as.data.frame(pred))

# Classification: classification "type" + PI
pred <- get_predicted(x, predict = "classification", ci = 0.95)
head(as.data.frame(pred))