influenceDiag {glmxdiag}

Influence Diagnostic Measures

Description

Calculates or extracts some influence diagnostic measures such as DFbetas, Cook's distance and leverage.

Usage

influenceDiag(model, approx = TRUE)

Arguments

model

a model supported by glmxdiag.

approx

logical, if TRUE the function is faster but returns approximated results.

Details

Leverage is extracted from the models using hatvalues function of each model class.

The elements for GLMs and negative binomial regression models are always approximated as they are extracted from influence function.

The argument approx can be useful only with models of class vglm and betareg since influence method function is not defined and items are fully calculated. It estimates n models where observations are excluded one by one. When approx = T one step approximation is used; the tol parameter within every fitting model function is set high such that the fitting process stops at the first iteration hence results are not exact but they are supposed to be close.

For betabinomial models, cook's distance is replaced with the quantity (b - b(-i)) Var(b) (b - b(-i)))/p as suggested by glmtoolbox package..

The aim of the function is to group these diagnostic measures in one list; one should use influenceDiag to calculate the measures and use the output object inside functions DFbeta, cookDist and leverage in order to graphically visualize results.

Value

Returns a list of class "influence" with the following elements:

DFbeta

data frame containing dfbetas for all observations and variables.

cookDist

a vector containing cook's distances.

leverage

hat values, i.e. the diagonal of the hat matrix.

full.beta

coefficients of the full model.

family

family name

Author(s)

Giuseppe Reale

Examples

data("mtcars")
mod <- glm(mpg ~ cyl + hp + carb + wt, family = Gamma, data = mtcars)
inf <- influenceDiag(mod)
cookDist(inf)
leverage(inf)
DFbeta(inf)