MuMIn (version 1.9.5)

model.avg: Model averaging

Description

Model averaging based on an information criterion.

Usage

model.avg(object, ..., revised.var = TRUE)
## S3 method for class 'default':
model.avg(object, ..., beta = FALSE, rank = NULL, rank.args = NULL,
    revised.var = TRUE, dispersion = NULL, ct.args = NULL)
## S3 method for class 'model.selection':
model.avg(object, subset, fit = FALSE, ..., revised.var = TRUE)

Arguments

object

a fitted model object or a list of such objects, or a model.selection object. See Details.

...

for default method, more fitted model objects. Otherwise, arguments that are passed to the default method.

beta

logical, should standardized coefficients be returned?

rank

optionally, a custom rank function (information criterion) to use instead of AICc, e.g. BIC or QAIC, may be omitted if object is a model list returned by get.models or a model.se

rank.args

optional list of arguments for the rank function. If one is an expression, an x within it is substituted with a current model.

revised.var

logical, indicating whether to use revised formula for standard errors. See par.avg.

dispersion

the dispersion parameter for the family used. See summary.glm. This is used currently only with glm, is silently ignored otherwise.

ct.args

optional list of arguments to be passed to coefTable (besides dispersion).

subset

see subset method for model.selection object.

fit

if TRUE, the component models are fitted using get.models. See Details.

Value

• An object of class averaging is a list with components:
• summary
  a data.frame with log-likelihood, IC, Delta(IC) and Akaike weights for the component models.
• coef.shrinkage
  a vector of full model-averaged coefficients, see Note.
• coefArray
  an array of component models' coefficients, their standard errors, and degrees of freedom.
• term.codes
  names of the terms with numerical codes used in the summary.
• avg.model
  the model averaged parameters. A data.frame containing averaged coefficients, unconditional standard error, adjusted SE (if dfs are available) and z-values (coefficient and SE) and significance (assuming a normal error distribution).
• importance
  relative importance of the predictor variables (including interactions), calculated as a sum of the Akaike weights over all of the models in which the parameter of interest appears.
• term.names
  character vector giving names of all terms in the model.
• x, formula
  the model matrix and formula corresponding to the one that would be used in a single model. formula contains only the averaged coefficients.
• call
  the matched call.
• In addition, the object has following attributes:
• modelList
  a list of component model objects.
• beta
  logical, were standardized coefficients used?
• revised.var
  if TRUE, the standard errors were calculated with the revised formula (See par.avg).

encoding

utf-8

Details

model.avg may be used either with a list of models, or directly with a model.selection object (e.g. returned by dredge). In the latter case, the models from the model selection table are not evaluated unless the argument fit is set to TRUE or some additional arguments are present (such as rank or dispersion). This results in much faster calculation, but has certain drawbacks, because the fitted component model objects are not stored, and some methods (e.g. predict, fitted, model.matrix or vcov) would not be available with the returned object. Otherwise, get.models is called prior to averaging, and ...are passed to it.

For a list of model types that are accepted see list of supported models.

rank is found by a call to match.fun and typically is specified as a function or a symbol (e.g. a back-quoted name) or a character string specifying a function to be searched for from the environment of the call to lapply. rank must be a function able to accept model as a first argument and must always return a scalar.

Several standard methods for fitted model objects exist for class averaging, including summary, predict, coef, confint, formula, vcov. The coef method a accepts argument full, if set to TRUE the full model-averaged coefficients are returned, rather than subset-averaged ones. logLik returns a list of logLik objects for the component models.

References

Burnham, K. P. and Anderson, D. R (2002) Model selection and multimodel inference: a practical information-theoretic approach. 2nd ed. New York, Springer-Verlag.

See Also

See par.avg for more details of model averaged parameter calculation.

dredge, get.models AICc has examples of averaging models fitted by REML.

modavg in package AICcmodavg, and coef.glmulti in package glmulti also perform model averaging.

Examples

# Example from Burnham and Anderson (2002), page 100:
library(MuMIn)
data(Cement)
fm1 <- lm(y ~ ., data = Cement)
(ms1 <- dredge(fm1))

#models with delta.aicc < 4
summary(model.avg(ms1, subset = delta < 4))

#or as a 95\% confidence set:
avgmod.95p <- model.avg(ms1, cumsum(weight) <= .95)
confint(avgmod.95p)

# The same result, but re-fitting the models via 'get.models'
confset.95p <- get.models(ms1, cumsum(weight) <= .95)
model.avg(confset.95p)

# Force re-fitting the component models
model.avg(ms1, cumsum(weight) <= .95, fit = TRUE)
# Models are also fitted if additional arguments are given
model.avg(ms1, cumsum(weight) <= .95, rank = "AIC")

# using BIC (Schwarz's Bayesian criterion) to rank the models
BIC <- function(x) AIC(x, k = log(length(residuals(x))))
model.avg(confset.95p, rank = BIC)
# the same result, using AIC directly, with argument k
# 'x' in a quoted 'rank' argument is substituted with a model object
# (in this case it does not make much sense as the number of observations is
# common to all models)
model.avg(confset.95p, rank = AIC, rank.args = alist(k = log(length(residuals(x)))))