influridge: Influence of single observations on the choice of the penalty...

In khellton/influridge: Visualize Influence of Observations in Ridge Regression

Influence of single observations on the choice of the penalty parameter in ridge regression

Description

This function produces a curve plot showing the optimal tuning paramer when up- or downweighting each observation.

Usage

influridge(
  X,
  y,
  nw = 40,
  max.weight = 4,
  noExpand = 0,
  noShrink = 0,
  degreeFreedom = FALSE,
  control.list = list(factr = 1e-04)
)

Arguments

X	numerical input matrix, of dimension nobs x nvars; each row is an observation vector.
y	numerical output vector; response variable.
nw	scalar, number of grid points for curve.
max.weight	scalar, the maximum weight for the curve.
noExpand	scalar, number of expanders to be highlighted (less than 5).
noShrink	scalar, number of shrinkers to be highlighted (less than 5).
degreeFreedom	logical, should the degrees of freedom be plotted instead of the tuning parameter (default = FALSE).
control.list	list, a list of control parameters for the optim function. See 'Details' under graphics::optim.

Examples

p <- 5
n <- 20
sigma <- 1
beta <- rep(1, p)

## Simulating design matrix, X
set.seed(556)
X <- matrix(rnorm(n * p), n, p)

## Simulate outcome vector, Y
y <- X %*% beta + rnorm(n, 0, sigma)

## Plot curves (no highlighted shrinkers/expanders)
influridge(X, y)

## Adding a large positive residual to observation 10 creates an influential shrinker
y[10] <- y[10] + 3
influridge(X, y, noShrink = 1, nw = 20)

## Plot degrees of freedom
influridge(X, y, noShrink = 1, nw = 20, degreeFreedom = TRUE)

## Not run: 
## Make plot for Body Fat dataset
require(mfp)
data(bodyfat)
X <- bodyfat[, 6:17] # Omit non-continous age variable
y <- bodyfat$siri
n <- dim(X)[1]

X <- scale(X, center = FALSE) # Scale data
X <- cbind(rep(1, n), X) # Add intercept to design matrix

influridge(X, y, noShrink = 1, noExpand = 1, degreeFreedom = TRUE)

## End(Not run)

khellton/influridge documentation built on Dec. 23, 2024, 9:24 a.m.