Skip to contents

Extract and Predict from LHWR Models

Source: R/extractors.R
fitted.lhwr.Rd

Convenience S3 methods for fitted LHWR models. These methods make LHWR objects behave like ordinary R model objects for common downstream tasks. predict.lhwr() returns fitted values for the calibration data when newdata is omitted. For new observations, supply both the Euclidean covariates in newdata and their Lorentz-hyperboloid coordinates in Znew. Prediction is then obtained by fitting a local weighted Gaussian or binomial model at each new Lorentz location using only the original calibration data.

Usage

# S3 method for class 'lhwr'
fitted(object, ...)

# S3 method for class 'lhwr'
residuals(object, ...)

# S3 method for class 'lhwr'
coef(object, ...)

# S3 method for class 'lhwr'
predict(
  object,
  newdata = NULL,
  Znew = NULL,
  type = c("response", "link", "coefficients"),
  dMat = NULL,
  bw = NULL,
  kernel = NULL,
  adaptive = NULL,
  bandwidth_type = NULL,
  geometry = NULL,
  ...
)

# S3 method for class 'hgwr'
fitted(object, ...)

# S3 method for class 'hgwr'
residuals(object, ...)

# S3 method for class 'hgwr'
coef(object, ...)

# S3 method for class 'hgwr'
predict(
  object,
  newdata = NULL,
  Znew = NULL,
  type = c("response", "link", "coefficients"),
  dMat = NULL,
  bw = NULL,
  kernel = NULL,
  adaptive = NULL,
  bandwidth_type = NULL,
  geometry = NULL,
  ...
)

Arguments

object

An LHWR model object returned by lhwr().

...

Unused.

newdata

Optional data frame containing predictor variables for new observations. If omitted, predict() returns fitted values from the calibration data.

Znew

Optional m x 3 Lorentz-hyperboloid coordinate matrix for newdata. Required when newdata is supplied.

type

Prediction type. "response" returns fitted responses (probabilities for binomial LHWR), "link" returns linear predictors, and "coefficients" returns the local coefficient matrix used for prediction.

dMat

Optional m x n distance matrix from new observations to the original calibration observations. If supplied, it is used directly.

bw

Optional prediction bandwidth. By default the bandwidth stored in the fitted model is used.

kernel

Optional kernel name. By default the fitted model kernel is used.

adaptive

Optional logical bandwidth flag. By default the fitted model setting is used.

bandwidth_type

Optional bandwidth type. By default the fitted model setting is used.

geometry

Optional prediction geometry. By default the fitted model geometry is used.

Value

fitted() returns the fitted response vector; residuals() returns residuals; coef() returns the n x p calibration local coefficient matrix. predict() returns either a numeric response vector or an m x p matrix of local coefficients for newdata.

Details

Tangent-E prediction reuses the stored training reference point for every prediction batch. If an older object has no stored reference, it is reconstructed from object$Z, never from Znew. Identical mapped points have exactly zero distance; nearby distinct tangent coordinates are not rounded to zero by a pairwise-distance tolerance. With unchanged model settings, predicting the training rows reproduces the calibration fit up to floating-point error.

Examples

set.seed(11)
H <- simulate_H2(50, r_max = 3)
x <- rnorm(50)
y <- 1 + 2 * x + rnorm(50, sd = 0.3)
fit <- lhwr(y ~ x, data.frame(y, x), Z = H$Z, bw = 20)
head(fitted(fit))
head(residuals(fit))
head(coef(fit))

Znew <- simulate_H2(3, r_max = 3)$Z
newdata <- data.frame(x = c(-1, 0, 1))
predict(fit, newdata = newdata, Znew = Znew)