R2BayesX: predict.bayesx – R documentation

Become an expert in R — Interactive courses, Cheat Sheets, certificates and more!

predict.bayesx

Prediction from fitted BayesX objects

Description

Takes a fitted "bayesx" object returned from bayesx and produces predictions by refitting the initial model with weights set to zero for new observations.

Usage

## S3 method for class 'bayesx'
predict(object, newdata, model = NULL,
  type = c("response", "link", "terms", "model"),
  na.action = na.pass, digits = 5, ...)

Arguments

`object`	an object of class `"bayesx"` or `"bayesx.hpc"`.
`newdata`	a data frame or list containing the values of the model covariates at which predictions are required. If missing `newdata` is the `model.frame` of the provided model.
`model`	for which model should predictions be calculated, either an integer or a character, e.g. `model = "mcmc.model"`. Note that exactly one model must be selected within argument `model` to compute predicted values!
`type`	when `type = "response"`, the default, predictions on the scale of the response are returned, `"link"` returns the linear predictor. When `type = "terms"`, each component of the linear predictor is returned, but excludes any offset and intercept. If `type = "model"`, the full model returned from updating the initial model with weights, that is used for computing predictions, is returned.
`na.action`	function determining what should be done with missing values in `newdata`.
`digits`	predictions should usually be based on the new values provided in argument `newdata`. However, since this prediction method uses refitting of the model with weights, predictions for model terms need to be matched with the new observations. BayesX returns values with a lower precision than R, therefore argument `digits` is used to round values when `type = "terms"`, to find matching `newdata` pairs in the fitted objects returned from the refitted model and the new data. Note that this is a workaround and not 100% bulletproof. It is recommended to compute predictions for `type = "response"` or `type = "link"`.
`...`	not used.

Value

Depending on the specifications of argument type.

Note

This prediction method is based on refitting the initial model with weights, i.e., if new observations lie outside the domain of the respective covariate, the knot locations when using e.g. P-splines are calculated using the old and the new data. Hence, if there are large gaps between the old data domain and new observations, this could affect the overall fit of the estimated spline, i.e., compared to the initial model fit there will be smaller or larger differences depending on the newdata provided.

Author(s)

Nikolaus Umlauf, Thomas Kneib, Stefan Lang, Achim Zeileis.

Examples

## Not run: 
## generate some data
set.seed(121)
n <- 500

## regressors
dat <- data.frame(x = runif(n, -3, 3), z = runif(n, 0, 1),
  w = runif(n, 0, 3))

## generate response 
dat$y <- with(dat, 1.5 + sin(x) + z -3 * w + rnorm(n, sd = 0.6))

## estimate model
b <- bayesx(y ~ sx(x) + z + w, data = dat)

## create some data for which predictions are required
nd <- data.frame(x = seq(2, 5, length = 100), z = 1, w = 0)

## prediction model from refitting with weights
nd$fit <- predict(b, newdata = nd)
plot(fit ~ x, type = "l", data = nd)

## End(Not run)

R2BayesX

Estimate Structured Additive Regression Models with 'BayesX'

v1.1-1

GPL-2 | GPL-3

Authors

Nikolaus Umlauf [aut, cre], Thomas Kneib [aut], Stefan Lang [aut], Achim Zeileis [aut] (0000-0003-0918-3766)

Initial release

2017-12-05