FLXMRrobglm {flexmix}

FlexMix Driver for Robust Estimation of Generalized Linear Models

Description

This driver adds a noise component to the mixture model which can be used to model background noise in the data. See the Compstat paper Leisch (2008) cited below for details.

Usage

FLXMRrobglm(formula = . ~ ., family = c("gaussian", "poisson"),
            bgw = FALSE, ...)

Arguments

formula

A formula which is interpreted relative to the formula specified in the call to flexmix using update.formula. Default is to use the original flexmix model formula.

family

A character string naming a glm family function.

bgw

Logical, controls whether the parameters of the background component are fixed to multiples of location and scale of the complete data (the default), or estimated by EM with normal weights for the background (bgw = TRUE).

...

passed to FLXMRglm

Value

Returns an object of class FLXMRrobglm inheriting from FLXMRglm.

Note

The implementation of this model class is currently under development, and some methods like refit are still missing.

Author(s)

Friedrich Leisch and Bettina Gruen

References

Friedrich Leisch. Modelling background noise in finite mixtures of generalized linear regression models. In Paula Brito, editor, Compstat 2008–Proceedings in Computational Statistics, 385–396. Physica Verlag, Heidelberg, Germany, 2008.
Preprint available at http://epub.ub.uni-muenchen.de/6332/.

Examples

## Example from Compstat paper, see paper for detailed explanation:
data("NPreg", package = "flexmix")
DATA <- NPreg[, 1:2]
set.seed(3)
DATA2 <- rbind(DATA, cbind(x = -runif(3), yn = 50 + runif(3)))

## Estimation without (f2) and with (f3) background component
f2 <- flexmix(yn ~ x + I(x^2), data = DATA2, k = 2)
f3 <- flexmix(yn ~ x + I(x^2), data = DATA2, k = 3,
              model = FLXMRrobglm(), 
              control = list(minprior = 0))

## Predict on new data for plots
x <- seq(-5,15, by = .1)
y2 <- predict(f2, newdata = data.frame(x = x))
y3 <- predict(f3, newdata = data.frame(x = x))

## f2 was estimated without background component:
plot(yn ~ x, data = DATA2, pch = clusters(f2), col = clusters(f2))
lines(x, y2$Comp.1, col = 1)
lines(x, y2$Comp.2, col = 2)

## f3 is with background component:
plot(yn ~ x, data = DATA2, pch = 4 - clusters(f3),
     col = 4 - clusters(f3))
lines(x, y3$Comp.2, col = 2)
lines(x, y3$Comp.3, col = 1)