forward function

F-test based model effect selection for linear models.

Adaptation of existing methods based on AIC/BIC.

forward(model, alpha = 0.2, full = FALSE, force.in)
backward(model, alpha = 0.2, full = FALSE, hierarchy = TRUE, force.in)
stepWise(model, alpha.enter = 0.15, alpha.remove = 0.15, full = FALSE)
stepWiseBack(model, alpha.remove = 0.15, alpha.enter = 0.15, full = FALSE)
wideForward(formula, data, alpha = 0.2, force.in = NULL)

Arguments

• model: object class lm to select effects from.
• formula: formula specifying all possible effects.
• data: data.frame corresponding to formula.
• alpha: numeric p-value cut-off for inclusion/exclusion.
• full: logical indicating extended output of forward/backward selection.
• force.in: character vector indicating effects to keep in all models.
• alpha.enter: numeric p-value cut-off for inclusion.
• alpha.remove: numeric p-value cut-off for exclusion.
• hierarchy: logical indicating if hierarchy should be forced in backward selection.

Details

F-based versions of built in stepwise methods.

Returns

The final linear model after selection is returned.

Author(s)

Kristian Hovde Liland

Examples

set.seed(0)
data <- data.frame(y = rnorm(8),
                                   x = factor(c('a','a','a','a','b','b','b','b')),
                                   z = factor(c('a','a','b','b','a','a','b','b')))
mod <- lm(y ~ x + z, data=data)
forward(mod)
backward(mod)
stepWise(mod)
stepWiseBack(mod)

# Forward selection for wide matrices (large number of predictors)
set.seed(0)
mydata <- data.frame(y = rnorm(6), X = matrix(rnorm(60),6,10))
fs <- wideForward(y ~ ., mydata)
print(fs)