glmChange: Maximal First Differences for Generalized Linear Models

glmChangeR Documentation

Maximal First Differences for Generalized Linear Models

Description

For objects of class glm, it calculates the change in predicted responses, for maximal discrete changes in all covariates holding all other variables constant at typical values.

Usage

glmChange(
  obj,
  data,
  V = NULL,
  typical.dat = NULL,
  change.dat = NULL,
  diffchange = c("range", "sd", "unit"),
  outcome = c("diff", "maxdiff"),
  n = 1,
  catdiff = c("biggest", "all"),
  sim = FALSE,
  R = 1000,
  qtiles = c(0.025, 0.975),
  ci_anchors = FALSE,
  adjust = c("none", "shift", "trim")
)

Arguments

obj

A model object of class glm.

data

Data frame used to fit object.

V

An optional variance-covariance matrix for the coefficients, if NULL, will be obtained through a call to vcov.

typical.dat

Data frame with a single row containing values at which to hold variables constant when calculating first differences. These values will be passed to predict, so factors must take on a single value, but have all possible levels as their levels attribute.

change.dat

A named list of values over which the variables of interest will be changed. If NULL or partially specified, unspecified variables will use diffchange. If a categorical variable is specified in here, this overrides the catdiff argument and only the specified contrast will be generated.

diffchange

A string indicating the difference in predictor values to calculate the discrete change. range gives the difference between the minimum and maximum, sd gives plus and minus one-half standard deviation change around the median and unit gives a plus and minus one-half unit change around the median.

outcome

For quantitative variables, should the difference over the range of chosen values be calculated (the default) or should the maximum probability difference over the range be calculated. These will be the same for single-term quantitative variables, but could be different for multi-term variables, like splines and polynomials.

n

number of units of diffchange to move. Only active for unit or sd.

catdiff

String identifying how differences in factor variables is handled. Options are "all" in which case all pairwise differences are returned, or "biggest" in which case the biggest difference is returned.

sim

Logical indicating whether simulated confidence bounds on the difference should be calculated and presented.

R

Number of simulations to perform if sim is TRUE

qtiles

Quantiles to calculate if sim=TRUE.

ci_anchors

Logical indicating whether the confidence intervals for the fit_Low and fit_High values are returned.

adjust

String identifying how range should be changed if it goes out of the bounds of the observed data. Trimming will simply truncate the size of the change to make it fit in bounds. Shifting will shift the interval so both ends are in bounds. If the shifted interval is wider than the range of the data, the change will be truncated to the range of the data.

Details

The function calculates the changes in predicted responses for maximal discrete changes in the covariates, for objects of class glm. This function works with polynomials specified with the poly function. It also works with multiplicative interactions of the covariates by virtue of the fact that it holds all other variables at typical values. By default, typical values are the median for quantitative variables and the mode for factors. The way the function works with factors is a bit different. The function identifies the two most different levels of the factor and calculates the change in predictions for a change from the level with the smallest prediction to the level with the largest prediction.

Value

A list with the following elements:

diffs

A matrix of calculated first differences

minmax

A matrix of values that were used to calculate the predicted changes

Author(s)

Dave Armstrong

Examples


data(france)
left.mod <- glm(voteleft ~ male + age + retnat + 
	poly(lrself, 2), data=france, family=binomial)
typical.france <- data.frame(
	retnat = factor(1, levels=1:3, labels=levels(france$retnat)), 
	age = 35, stringsAsFactors=TRUE
	)
glmChange(left.mod, data=france, typical.dat=typical.france)


davidaarmstrong/damisc documentation built on Oct. 1, 2023, 3:05 p.m.