augment.rq: Augment data with information from a(n) rq object

Description

Augment accepts a model object and a dataset and adds information about each observation in the dataset. Most commonly, this includes predicted values in the .fitted column, residuals in the .resid column, and standard errors for the fitted values in a .se.fit column. New columns always begin with a . prefix to avoid overwriting columns in the original dataset.

Users may pass data to augment via either the data argument or the newdata argument. If the user passes data to the data argument, it must be exactly the data that was used to fit the model object. Pass datasets to newdata to augment data that was not used during model fitting. This still requires that at least all predictor variable columns used to fit the model are present. If the original outcome variable used to fit the model is not included in newdata, then no .resid column will be included in the output.

Augment will often behave differently depending on whether data or newdata is given. This is because there is often information associated with training observations (such as influences or related) measures that is not meaningfully defined for new observations.

For convenience, many augment methods provide default data arguments, so that augment(fit) will return the augmented training data. In these cases, augment tries to reconstruct the original data based on the model object with varying degrees of success.

The augmented dataset is always returned as a tibble::tibble with the same number of rows as the passed dataset. This means that the passed data must be coercible to a tibble. If a predictor enters the model as part of a matrix of covariates, such as when the model formula uses splines::ns(), stats::poly(), or survival::Surv(), it is represented as a matrix column.

We are in the process of defining behaviors for models fit with various na.action arguments, but make no guarantees about behavior when data is missing at this time.

Usage

# S3 method for rq
augment(x, data = model.frame(x), newdata = NULL, ...)

Value

A tibble::tibble() with columns:

.fitted: Fitted or predicted value.
.resid: The difference between observed and fitted values.
.tau: Quantile.

Arguments

x

An rq object returned from quantreg::rq().

data

A base::data.frame or tibble::tibble() containing the original data that was used to produce the object x. Defaults to stats::model.frame(x) so that augment(my_fit) returns the augmented original data. Do not pass new data to the data argument. Augment will report information such as influence and cooks distance for data passed to the data argument. These measures are only defined for the original training data.

newdata

A base::data.frame() or tibble::tibble() containing all the original predictors used to create x. Defaults to NULL, indicating that nothing has been passed to newdata. If newdata is specified, the data argument will be ignored.

...

Arguments passed on to quantreg::predict.rq

object: object of class rq or rqs or rq.process produced by rq

interval

type of interval desired: default is 'none', when set to 'confidence' the function returns a matrix predictions with point predictions for each of the 'newdata' points as well as lower and upper confidence limits.

level

converage probability for the 'confidence' intervals.

type

For predict.rq, the method for 'confidence' intervals, if desired. If 'percentile' then one of the bootstrap methods is used to generate percentile intervals for each prediction, if 'direct' then a version of the Portnoy and Zhou (1998) method is used, and otherwise an estimated covariance matrix for the parameter estimates is used. Further arguments to determine the choice of bootstrap method or covariance matrix estimate can be passed via the ... argument. For predict.rqs and predict.rq.process when stepfun = TRUE, type is "Qhat", "Fhat" or "fhat" depending on whether the user would like to have estimates of the conditional quantile, distribution or density functions respectively. As noted below the two former estimates can be monotonized with the function rearrange. When the "fhat" option is invoked, a list of conditional density functions is returned based on Silverman's adaptive kernel method as implemented in akj and approxfun.

na.action

function determining what should be done with missing values in 'newdata'. The default is to predict 'NA'.

Details

Depending on the arguments passed on to predict.rq via ..., a confidence interval is also calculated on the fitted values resulting in columns .lower and .upper. Does not provide confidence intervals when data is specified via the newdata argument.

Examples

Run this code


# load modeling library and data
library(quantreg)

data(stackloss)

# median (l1) regression fit for the stackloss data. 
mod1 <- rq(stack.loss ~ stack.x, .5)  

# weighted sample median
mod2 <- rq(rnorm(50) ~ 1, weights = runif(50))

# summarize model fit with tidiers
tidy(mod1)
glance(mod1)
augment(mod1)

tidy(mod2)
glance(mod2)
augment(mod2)

# varying tau to generate an rqs object
mod3 <- rq(stack.loss ~ stack.x, tau = c(.25, .5))

tidy(mod3)
augment(mod3)

# glance cannot handle rqs objects like `mod3`--use a purrr 
# `map`-based workflow instead