These functions, applied on a glmulti object, produce model-averaged estimates, unconditional confidence intervals, and predictions
from the models in the confidence set (or a subset of them). They are equivalents of the standard coef and predict for single models.
# S3 coef method for class 'glmulti'
# S3 method for glmulti
coef(object, select="all", varweighting="Buckland",
icmethod="Lukacs", alphaIC=0.05, ...)# S3 predict method for class 'glmulti'
# S3 method for glmulti
predict(object, select="all", newdata=NA, se.fit=FALSE,
varweighting="Buckland", icmethod="Lukacs", alphaIC=0.05, ...)
an object of class glmulti
A specification of which models should be used for inference. By default all models are used, see below.
The method to be used to compute the unconditional variance. "Buckland" (the default) (implements the approach presented in Buckland et al. 1997. "Johnson" implements a slightly different approach recommended in Johnson \& Omland 2004 and proposed at page 235 in Burnham \& Anderson 2002. The latter results in slightly bigger estimates of the unconditional variance of model coefficients.
Method to construct confidence intervals. One of "Standard", "Burnham" or "Lukacs". The three methods differ in their use (or not) of degrees of freedom.
New data.frame of data for which to predict values
Whether to return unconditional variances and confidence intervals associated with predicted values
The alpha risk when building the confidence intervals
Further arguments to single-model coef or predict
coef returns a data.frame with model-averaged estimates of the different parameters in the models, as well as their unconditional variance, importance, and confidence interval according to one of three methods: "Standard" simply assumes a Normal distribution of the estimator (Buckland 1997), "Lukacs" assumes a Student distribution with degrees of freedom taken to be averaged across models (see Lukacs et al. 2010), and "Burnham" is a more sophisticated Student-based method proposed by Burnham \& Anderson 2002.
predict returns a list of three elements: the multi-model predictions, their variability (unconditional variance and confidence interval, if se.fit=T), and the number of NA predicted values that were treated as zeros when averaging models.
select can be used to specify which models should be used for inference. By default all are used. If specifying an integer value x, only the x best models are used. If a numeric value is provided, if it less than one, models that sum up to x% of evidence weight are used. If it more than one, models within x IC units from the best model are used.
Buckland et al. 1997. Model selection: an integral part of inference. Biometrics. Burnham \& Anderson. 2002. Model Selection and Multimodel Inference. An Information Theoretic Practical Approach. Johnson \& Omland. 2004. Model selection in ecology and evolution. TREE. Lukacs et al. 2010..Model selection bias and Freedman's paradox. Annals of the Institute of Statistical Mathematics.