Tools for performing model selection and model averaging. Automated model selection through subsetting the maximum model, with optional constraints for model inclusion. Model parameter and prediction averaging based on model weights derived from information criteria (AICc and alike) or custom model weighting schemes.

MuMIn package news:

Version: 1.43.6 (2019-04-08) (fixed) if conditions with "length > 1 in coercion to logical" error produced by the new Rcheck. (fixed) updated 'tests' to make up for 'spautolm' moved from 'spdep' to 'spatialreg'.

Version: 1.43.4 (2019-04-04) (fixed) minor changes to the quasi-likelihood function

Version: 1.43.3 (2019-03-19) (added) methods for 'gamlss' (suggested by Anne Loosen) (fixed) QIC,quasiLik: now gives results compatible with those of wgeesel::QIC.gee.

Version: 1.43.2 (2019-03-16) 'sw' (for "Sum of Weights") is now preferred to 'importance' (which is still available, but is going to be deprecated in future). This is because of the controversies about SoW being a valid measure of actual variable importance. [p]dredge: max. number of predictors is now correctly calculated (depending on the number of variants)

Version: 1.43.1 (2019-03-06) (added) new model structure for unmarkedFitGDS (gdistsamp) with alpha parameter (thanks to Alba Estrada)

Version: 1.43.0 (2019-02-28) (fixed) in QICu, the number of parameters was mistaken for number of observations (reported by Richard Fredrickson) (added) some methods for wgeesel::wgee objects.

Version: 1.42.5 (2019-01-02)
r.squaredGLMM: added the argument "envir" which is passed to `.nullFitRE`

. This
can be used in case of issues with model data not being found (reported by
Eliot McIntire)

Version: 1.42.4 (2018-12-05) (fixed) r.squaredGLMM.merMod: error when the function was called in an environment different than .GlobalEnv (spotted by a certain Martin on Stackoverflow)

Version: 1.42.3 (2018-07-31)
(added) dredge: the subset expression `.(x)`

now accepts an additional argument
for the allowed term orders. For example `.(X, 2:10)`

means
"at least one 2nd or 3rd order interaction of 'X'". The `.`

pseudo-function
has also a new alias: "with", e.g. `with(X, 2:10)`

.

Version: 1.42.2 (2018-07-27) (fixed) r.squaredGLMM: handle the case when some coefficients are not estimated and NA

Version: 1.42.1 (2018-07-21) tests have been modified to prevent CRAN check errors on some systems. (fixed) arm.glm, armWeights: now use a random sample to select a training set. armWeights: no longer give a warning about "longer object length"

Version: 1.42.0 (2018-07-17)
(updated) methods for "unmarkedFit" models have been rewritten and should work
properly with all subclasses. Added missing support for
"unmarked::unmarkedFitOccuFP" (reported by Tara Hohoff)
R2GLMM added an option to disable "null fit warning"
(fixed) getAllTerms.lme: now gives a correct value for response in "binomial"
glmmPQL ("zz" previously)
(fixed) testSmoothKConsistency: (used by `model.avg`

with 'gam' models) works
now with 'gam's including no smooth terms.

Version: 1.41.1 (2018-07-13)
(fixed) workaround for clmm's (package "ordinal") annoyance: it needs explicit
"1" in the formula if there are no other fixed terms. That resulted in
intercept-only model calls generated by `dredge`

being rejected.
(kindly reported by J�rg Albrecht)

Version: 1.41.0 (2018-07-12) (updated) r.squaredGLMM: revised to incorporate new developments in Nakagawa et al (2017). Methods exist for merMod, glmmTMB, glmmADMB, glmmPQL, cpglmm, and lm classes.

Version: 1.40.8 (2018-06-18)
Model description - simplified handling of the model's family description in
the model selection tables, so it does not require `family()$family`

string
being an existing function name. The drawback is that now a default link
function is not detected and its name always appears in the column (
before e.g. "binomial(logit)" appeared as "binomial()").

Version: 1.40.7 (2018-06-14) Addad a workaround to accommodate glmmTMB specific families (which do not have a matching function) (reported by a certain Gitu on Stackoverflow)

Version: 1.40.6 (2018-05-16) (fixed) makeArgs, getAllTerms: compatibility with glmmTMB - random terms in formulas other than cond have not been handled properly (discovered via Mollie Brooks on GitHub)

Version: 1.40.5 (2018-05-14)
`.checkNAAction`

now handles a case of 'na.action' formal argument being a call
(e.g. `getOption(na.action)`

in glmmTMB) (spotted by Ben Bolker)

Version: 1.40.4 (2018-01-30) Elaborated the Description field.

Version: 1.40.3 (2018-01-29)
Removed superfluous method replacements for models from package "survival".
`predict`

functions for "gls" and "lme" are no longer registered as S3
methods in NAMESPACE.

Version: 1.40.2 (2018-01-19) (fixed) getAllTerms.glmmTMB: error with models including offset term (spotted by Kate Plummer)

Version: 1.40.1 (2017-11-30) (added) Support for "glmmTMB" models (incentive from Kaitlyn Zerr)

Version: 1.40.0 (2017-09-30) (fixed) long broken support for "unmarkedFit" models has been restored (because of the utterly inconsistent structure of these objects it is not guaranteed to work in all cases). (fixed) R.squaredGLMM: error with very long RE formulas (reported by Siti Binte Faizal)

Version: 1.16.7 (2017-08-10) A few minor things fixed.

Version: 1.16.6 (2017-07-11) (fixed) print.summary.averaging: warning when fitting function name was a call rather than a symbol, e.g. lme4::lme (reported by Chen Lei).

Version: 1.16.5 (2017-05-24) (fixed) predict.gls updated to match current version in nlme (reported by a certain Asteraceae on Stackoverflow)

Version: 1.16.4 (2016-11-21) (fixed) makeArgs.mark: reverted to a working version (this bug was introduced in 1.15.2/rev. 373. Reported by Jay Rotella)

Version: 1.16.3 (2016-09-14) (fixed) dredge: recognizes when global model's data is a function in a smarter way

Version: 1.16.2 (2016-09-01) (fixed) plot.model.selection: problem in case with only one model in the table (reported by Ryan Rothman on Stackoverflow) dredge: recognizes when global model's data is a function (which is not allowed as it may cause sub-models being fitted to different datasets)

Version: 1.16.1 (2016-07-29)
New suite of functions to support model averaging. It is based on original
code developed by Carsten Dormann, as an outcome of the model-averaging
workshop held in Freiburg in 2015.
(added) cos2Weights, BGWeights, jackknifeWeights, stackingWeights: calculate
various types of model weights.
(added) loo: RMSE/log-likelihood based on leave-one-out cross-validation
(added) armWeights: similar to `arm.glm`

but produces model weights.
(fixed) model.avg/model.sel: problems with intercept only models
(spotted by a certain Teresa on Stackoverflow)

Version: 1.15.8 (2016-07-05) (fixed) dredge: no longer produces an error in the rare case when dependency matrix was an empty matrix [0x0]. (fixed) r.squaredGLMM: error with naming of variance components in 'lmer' models. (both problems reported by a certain Teresa on Stackoverflow) (added) Weights: has an assignment version, so weights can be applied to an "averaging" object. get.models: now can extract models also from an "averaging" object (only when it was created with model objects).

Version: 1.15.7 (2016-01-21) (fixed) dredge: number of variables is limited to 30 to avoid integer overflow (reported by Sarah Grogan).

Version: 1.15.6 (2015-12-21)
(fixed) get.models, dredge, subset.model.selection: enclosure for the subset
expression evaluation is now `parent.frame()`

rather than `.GlobalEnv`

as it
was before (incentive from Philipp Maier).
(fixed) coeffs.survreg, coefTable.survreg: log(scale) added to coefficients,
when it is estimated (suggested by Nicolas Rode).

Version: 1.15.5 (2015-11-25) (fixed) [p]dredge: error with model with offsets in formula when 'subset' expression was given (reported by Adam Cummings) (fixed) [p]dredge: in some cases the generated combinations did not obey the subset matrix (reported by Ben Augustine) (fixed) model.selection column assignment "[<-" was broken when (spotted by a certain abmiller8 on Stackoverflow) model.avg: warns when averaged 'gam' models have smooths with different dimension for the same variable.

Version: 1.15.4 (2015-10-16) (fixed) rbind, merge methods for "model.selection": wrong order of names of columns and rows (spotted by Ben Augustine)

Version: 1.15.3 (2015-09-06) (fixed) Weights: bug that caused that NaN's were returned (reported by Carles Alcaraz)

Version: 1.15.2 (2015-07-07) (added) summary.coefTable: tries to calculate p-values for coefficients, nicer printing using printCoefmat

Version: 1.15.1 (2015-07-02) added imports from all the default packages added to NAMESPACE in order to satisfy CRAN's new stricter RCheck.

Version: 1.50.0 (2015-07-01) (added) arm.glm: implements model averaging with weights calculated via ARM ("adaptive regression by mixing") algorithm. "averaging" object: the elements "coef.shrinkage" and "coefTable" are replaced by matrix named "coefficients". The "Pval-tables" are accessed via summary(averaging)'s elements "coefmat.full" and "coefmat.subset". Objects created with previous versions will be automatically converted if needed. [p]dredge: argument 'm.lim' replaces 'm.min' and 'm.max' (which are still present, but deprecated). "model.selection" object: changing or removing essential columns (i.e. everything except 'varying' and 'extra') converts the object to a plain "data.frame" and drops all additional attributes. (added) rbind.model.selection: combine multiple "model.selection" objects. (added) 'xtable' methods for "averaging" and "model.selection" objects. (fixed) updateable: does not cause infinite recursion when the resulting wrapper has the same name as the wrapped function. (fixed) exprApply: now can deal with function expressions and 'expression's longer than one.

Version: 1.14.1 (2015-06-12) (fixed) [p]dredge no longer gives coefficient values as NA with [g]lmer (reported by Aitor Cevidanes). Some tidying-up of manual and data (response in 'Cement' and 'Beetle' moved to first column so that formula(data) is correct).

Version: 1.14.0 (2015-06-03) (added) partial.sd, std.coef: calculate partial std. deviation and standardize model coefficients. 'beta.weights' is now deprecated in favour of 'std.coef'. dredge, model.avg, model.sel: argument 'beta' is modified to allow for standardizing with SD or partial SD. coefTable, coeffs: method for class 'aodml' (package 'aod3') no longer return 'phi' coefficients. This is for consistency with results of 'vcov' and 'coef', as well as their counterparts for 'glimML' models. Removed all methods for 'mer' objects (used in lme4 < 1.0.0)

Version: 1.13.10 (2015-06-01) (fixed) compatibility with lme4-1.1.8, 'null.fit' now works properly with 'gamm' objects.

Version: 1.13.9 (2015-05-28) stdize: improved interpretation of the argument "scale". new dataset "GPA" ("First-year college Grade Point Average" referred in Burnham & Anderson (2002)).

Version: 1.13.8 (2015-05-22) (added) QIC support for 'geem' models (request from Lee McDaniel).

Version: 1.13.7 (2015-05-14) coefTable: method for 'averaging' objects gains new arguments "full" and "adjust.se". (fixed) getAllTerms: now correctly names coefficients of 'unmarked::gdistsamp' models (reported by Nathan Hostetter)

Version: 1.13.6 (2015-05-01) (fixed) model.avg, model.sel: now better distinguish model objects from lists of objects.

Version: 1.13.5 (2015-04-15) (fixed) r.squaredGLMM: method for 'lme' class with a multilevel grouping structure (reported by Alex Gunderson)

Version: 1.13.4 (2015-02-24) coefTable.coxph: added workaround for the case when variance-covariance matrix is NA (as in results of 'intcox') (reported by Vanessa Boukili)

Version: 1.13.3 (2015-02-19) get.models+subset.model.selection: unified 'subset' expression interpretation dredge: no longer exceeds memory limit with very complex models, and its efficiency has somewhat improved. With trace=2 a progress bar is displayed. model.sel: has new attribute 'fit'. "averaging" object now stores component models. model.sel,subset,get.models: unified interpretation of the "subset" expression (fixed) extracting random effects formula (affected 'r.squaredGLMM' for 'lmer', reported by Yevgen Matusevych) (fixed) makeArgs (and hence 'dredge'): formulas with no intercept are now handled correctly in 'zeroinfl' models.

Version: 1.13.2 (2015-01-08) (fixed) makeArgs (and hence 'dredge'): gave error if no random effects was specified in 'glmmadmb' model (reported by Szymek Drobniak).

Version: 1.13.1 (2015-01-06) (fixed) getAllTerms: now properly handles single formula in 'hurdle' and 'zeroinfl' models (package 'pscl').

Version: 1.13.0 (2015-01-03) (added) get.response: utility to extract response variable.

Version: 1.12.4 (2014-12-26) (added) stdizeFit: utility to modify model object or a call to modelling function to use standardized data.

Version: 1.12.3 (2014-12-23) (fixed) checkModels: workaround for the "dotted argument identity" issue (namely sometimes equal "..1"s are not identical).

Version: 1.12.2 (2014-12-19) (fixed) [Q]AICc: previously they ignored REML argument for some model classes. The code has been rewritten so that most of the core calculation is done by a common internal function.

Version: 1.12.1 (2014-12-16) minor corrections to the documentation

Version: 1.12.0 (2014-12-16) (added) exprApply: new function to manipulate expressions (with example of use with 'dredge' to manipulate terms inside formula). dredge, subset.model.selection: curly braces can be used to indicate complex terms (e.g. {I(x)} or {s(x,k=2}) instead of backtick quotes. Spacing in these expressions is irrelevant (as both whitespace and backticks seem to be a problem for some users).

Version: 1.11.3 (2014-12-10) (added) stdize: new function for standardizing data (fixed) smarter sorting of interaction terms in coefficient names print.model.selection: rank function call is added to the output

Version: 1.11.2 (2014-12-01) (fixed) r.squaredGLMM: in "binomial" family models the residual variance is used for additive dispersion (thanks to Paul Johnson). There is still issue with 'lmer' with 'subset' argument, and observation-level random term has to be specified manually. (fixed) some formula-decomposition related functions (that sometimes caused problems with 'r.squaredGLMM' for 'lmer', reported by Yevgen Matusevych) dredge: warns about potentially misspecified arguments (added) predict method for 'gamm' class

Version: 1.11.1 (2014-11-19) dredge: more efficient checking for marginality in model terms (~10x faster in case of 4-way interaction). The argument 'marg.ex' is removed (exceptions to marginality rule are inferred from the global model). getAllTerms: result has a new attribute 'deps' with a "term dependency matrix". nested: new function listing nested models (suggested by Giacomo Assandri). updateable: added new argument 'eval.args'

Version: 1.10.8 (2014-11-14) dredge, subset.model.selection: improved interpretation of 'subset' expressions. "Term(x)" notation can be used instead of backtick quotes for complex term names.

Version: 1.10.7 (2014-09-24) (fixed) coef.multinom: gave a wrong order of coefficients for multiple level responses (thanks to Alistair Senior who spotted this)

Version: 1.10.6 (2014-08-17) (added) methods for 'maxlike' (from package with the same name) lazy loading of datasets plot.model.selection: improved and documented

Version: 1.10.5 (2014-08-08) get.models: can now use parallel computation if a valid 'cluster' is provided. It takes over the functionality of 'pget.models', which is now deprecated. Also, the argument 'subset' must be now explicitely provided (so all models are no longer fitted by default).

Version: 1.10.4 (2014-08-02) (fixed) importance, model.avg: wrong order of 'N containing models' attribute (bug reported by Alba Estrada) r.squaredGLMM: should be considerably faster and memory efficient due to improved calculation of random effects variance (slow performance with large datasets reported by Daniel Ezra Johnson).

Version: 1.10.3 (2014-06-27) (fixed) dredge: no longer drops random term in 'glmmadmb' models (spotted by Carlos Le�n)

Version: 1.10.2 mystery update

Version: 1.10.1 (2014-05-19) dredge: more meaningful warning message about misspecified 'fixed' argument pdredge: synchronized code with 'dredge' documentation: improved formatting

Version: 1.10.0 (2014-05-02) r.squaredGLMM: throws an error if contrasts that have been used for random factors in the "merMod" model do not match the ones in model matrix. This happens if there are factors in random slope formula and user has changed relevant 'options(contrasts)' since the model was fitted. dredge: 'na.action' checking now allows "option(na.action)" to be a function and not only character string.

Version: 1.9.26 (2014-03-16) (fixed) dredge: when columns/rows in a subset matrix were in different order than model terms, the values became wrongly rearranged (spotted by a certain Frank on Stackoverflow).

Version: 1.9.25 (2014-03-13) (added) methods for 'cpglm' and 'cpglmm' (from package 'cplm') (suggested by Andrew Cooper)

Version: 1.9.24 (2014-02-26) (fixed) 'getAllTerms.unmarkedFit' gave error when 'alpha' was estimated in pcount model (reported by Lindsey Garner)

Version: 1.9.23 (2014-02-24) (added) methods 'model.matrix' and 'model.frame' for 'lme', because the default ones give wrong results when there is 'subset' or 'contrasts' set, or 'na.omit'ting.

Version: 1.9.22 (2014-02-20) r.squaredGLMM: can be now applied to models with random slopes (based on idea and R code by Paul Johnson). Added calculation of conditional R^2 for Poisson family models.

Version: 1.9.21 (2014-02-17) logLik.asreml: number of estimated parameters ('df') is now calculated as the number of fixed effect coefficients + number of variance parameters (non-fixed and non-constained). This gives comparable numbers to similar lmer models. In 'Asreml-R manual', however, only the number of variance components is used as K for AIC calculation (page 15). The log-likelihood values of asreml are far different from those from lmer(REML = TRUE), even though coefficients are very close. Note also that REstricted-likelihood is returned, so model comparisions based on it are invalid. r.squaredGLMM: checks if it is applied to a random intercepts model, otherwise error is given. This is a limitation of the original formula by Nakagawa & Schiezlich.

Version: 1.9.20 (2014-02-12) (added) methods for 'phylolm' (suggestion by Rafael Maia)

Version: 1.9.19 (2013-12-21) (added) 'merge.model.selection' combines two 'model.selection' tables. Note that this is different than 'merge' for 'data.frame's (requested by Rafael Maia)

Version: 1.9.18 (2013-12-18) (added) methods for 'asreml' model (in commercial package 'asreml'). Note this has not been tested extensively (thanks to Emeline Mourocq for an incentive and testing)

Version: 1.9.17 (2013-12-14) full-averaged coefficients and standard errors are now calculated by 'model.avg' and relevant methods (argument full = TRUE) (suggested by Candan Soykan) importance: result gives now also number of models containg each variable

Version: 1.9.16 (2013-11-24) model.sel: now accepts arguments 'extra' and 'beta', like 'dredge' (requested by D. Duro on R-Forge).

Version: 1.9.15 (2013-11-17) [p]dredge: improved checking for whether the global model removes missing data (na.action = "na.omit" or "na.exclude"). Note that code working with previous versions may give you an error now, setting options(na.action = "na.fail") prevents it.

Version: 1.9.14 (2013-11-03) (added) support for 'betareg' (from package 'betareg') (suggested by Marie-Line Gentes) (added) support for 'aodml' and 'aodql' (from package 'aods3') (suggested by Renaud Lancelot) [p]dredge: fixed a bug that occurred when there was 'na.action=na.omit' in global model (spotted by Laura Busia on R-help).

Version: 1.9.13 (2013-10-28) (fixed) handling response variables with more than two levels in 'nnet::multinom' models (reported by Ludo Duvaux on Stackoverflow)

Version: 1.9.12 (2013-10-21) further adjustments to deal with unavailability of 'lme4' on CRAN

Version: 1.9.11 (2013-09-30) Documentation: examples that depend on suggested packages have been made conditional on availability of the package getCall method for 'model.selection' can retrieve a component model call if index is given as a second argument.

Version: 1.9.10 (2013-08-27) pdredge: removed global assignments (within cluster nodes).

Version: 1.9.9 (2013-05-21) (fixed) beta.weights: bug with dimensions dropped (spotted by Dennis Duro, thanks to Ben Bolker for tracking it down)

Version: 1.9.8 (2013-05-02) (fixed) handling bar in formulas for 'hurdle' class of models (reported by Jean-Yves Barnagaud)

Version: 1.9.7 (2013-04-19) dredge: 'subset' expression now recognizes ".(x)" notation to indicate a variable rather than a model term. Error message from within a 'rank' function gives now a more meaningfull call (apparently it is a common mistake to have some misspelled arguments that get passed to 'rank' function through '...').

Version: 1.9.6 (2013-04-12) model.sel: better model naming in the result table (fixed) logLik.coxme: 'nobs' attribute now uses a correct number of observations (model$n[2]). "penalized" log-lik. is the default type now (Note that "logLik.coxme"'s argument "type" is undocumented).

Version: 1.9.5 (2013-04-04) (fixed) AICc: now takes number of observations from "nall" (if it is available) attribute of the 'logLik' object. If not present, "nobs" attribute is used. r.squaredGLMM: current implementation does not calculate the marginal variant correctly for Poisson family models, so until it is resolved 'NA' is returned, with a warning.

Version: 1.9.4 (2013-04-01) (fixed) pdredge: bug with wrong number of updated rows. Replaced all calls to the soon-to-be deprecated ".find.package" with "find.package".

Version: 1.9.3 (2013-02-21) (fixed) method registration for 'getAllTerms.gam' was missing (reported by Steven Delean).

Version: 1.9.2 (2013-02-01) updateable: new function to create a wrapper for functions returning non-updateable objects (adds a 'call' element to the function's result). (renamed) removed 'gamm' wrapper, which is now replaced by 'uGamm'. (added) 'predict' method replacement for 'gls' models (adds 'se.fit' argument). model.avg: no longer returns residuals (this is because residuals for models other than 'lm' can be calculated in several different ways - with different results).

Version: 1.9.1 (2013-01-24) (fixed) bugs with 'MCMCglmm' models and 'varying' lists in 'dredge'.

Version: 1.9.0 (2013-01-23)
This version brings in more elaborated subsetting of models in 'dredge' and
'model.selection' object, and a new variant of pseudo-R^2.
[p]dredge: 'subset' can use 'varying' variables (as "V(varying.variable)") and
number of model terms (as `*nvar*`

). This allows for better fine-tuning of
the selection table.
'fixed' argument can be set to TRUE to fix all terms of 'global.model'.
[p]dredge, subset.model.selection: 'subset' expressions can now make use of a
new function 'dc' ("dependency chain") to make inclusion of variables
conditional on other variables.
r.squaredGLMM: new function to calculate R^2 statistic for mixed effect models,
described by Nakagawa & Schiezlich (2012).
(added) support for 'splm' (from package 'splm'), and 'logistf' models (from
package 'logistf').
Cp: now accepts more than one model (like 'AIC').
New demos "dredge.varying", "dredge.subset" and "gees".

Version: 1.8.2 (2012-12-07) (added) support for models of class 'mark' (from package 'RMark'). 'dredge' can currently only manipulate 'formula' element of argument 'model.parameters'. simplify.formula, expand.formula: new functions to convert formula between expanded and shorthand notation (i.e. between 'a+b+a:b' and 'a*b'). model.names: has new argument "use.letters".

Version: 1.8.1 (2012-11-23) (added) methods for class 'merMod' to accommodate recent changes in lme4_0.99999911-0. getAllTerms: now recognizes random effects in 'MCMCglmm'.

Version: 1.8.0 (2012-11-21) This version introduces support for GEE models ('gee','geeglm' and 'yags') ranked by QIC (note that 'yags' seems to be still quite buggy). Also 'clmm' and 'clm' from package 'ordinal' are now handled properly. QIC, quasiLik: new functions for use with GEE models. [p]dredge: argument 'subset' can be now a logical matrix representing allowed parameter combinations. Some improvements for speed have been made. [p]dredge, model.avg: new argument 'ct.args' for passing extra arguments to 'coefTable'.

Version: 1.7.12 (2012-11-11) tests/classes.R: fixed compatibility spdep-0.5.53. (spdep::spautolm now uses method = "eigen", was "full")

Version: 1.7.11 (2012-07-13) coefTable.multinom: fixed an error when coefficients and standard errors were matrices. get.models: now checks for the object class of the passed arguments.

Version: 1.7.10 (2012-06-07) (fixed) dredge: bug with an error message when number of predictor exceeded maximum (thanks to Corina Logan for spotting this). manual: added 'glmmADMB' to the list of supported classes (reported by Rocio Jana). 'coxme' moved from 'Suggested' to 'Enhances' (to avoid build error when the dependency is missing).

Version: 1.7.9 (2012-04-03) (added) 'predict' methods for 'lme' and 'lmer' (both can calculate 'se.fit'). 'predict.lme' is a wrapper for its namesake from 'nlme' package (it is not exported, so it's visibility depends on the order the packages were loaded - but is always visible from MuMIn namespace).

Version: 1.7.8 (2012-04-01) (fixed) 'coef.model.selection' returns now proper coefficients

Version: 1.7.7 (2012-04-01) (fixed) dependencies in tests

Version: 1.7.6 (2012-04-01) dredge: with nested formula designs, 'marg.ex' is now found automatically predict.averaging: is now more flexible in how the predictions from generalized models are calculated (new attribute "backtransform").

Version: 1.7.5 (2012-03-27) (added) methods for 'MCMCglmm' class, and function 'DIC': works with 'model.sel' and 'dredge' (but not 'model.avg'). 'unmarked' is removed from 'Suggested' packages (recently it is often unavailable on CRAN and causes errors during R CHECK). Added workarounds to pass (=cheat) the R CHECK.

Version: 1.7.3 (2012-02-01) (added) methods for 'glimML' class from package 'aod'

Version: 1.7.2 (2012-01-30) (fixed) 'nobs.coxme' returned the number of events was returned rather than the number of observations. (fixed) 'coefTable' gave an error for 'coxph' null model. (fixed) 'r.squaredLR' was evaluating the null model for 'lme' in a wrong environment. (fixed) 'getAllTerms' method for 'zeroinfl' was not registered. dredge: checks each generated model for the correct number of observations.

Version: 1.7.1 (2012-01-26) (added) methods for 'hurdle' and 'zeroinfl' classes (package 'pscf'). coef, coefTable: new methods to extract coefficients from model selection tables. Documentation: list of supported models put into a separate page.

Version: 1.7.0 (2012-01-25) This version adds possibility to model average and re-rank (with another IC) a model selection table without recreating the fitted models. Models from package 'coxme' are now supported. Model selection tables include now extra columns showing differences in attributes of the modelling function call. model.avg: use 'model.selection' without refitting the models. Fixed a problem with averaging different types of models together (happened when 'df's were available only for some of component models). dredge: coefficient tables added to the result (attribute "coefTables"). The code has been simplified. print.model.selection: now reports random terms also for objects generated by 'model.sel'. get.models: can handle 'model.selection' objects from 'model.sel' model.sel: re-ranking 'model.selection' without recreating original models (works only if the IC can be applied to logLik). Output has additional columns showing differences between models (based on call's arguments) (added) generic methods for 'coxme' and 'lmekin' classes (package 'coxme').

Version: 1.6.6 (2012-01-07) model.avg: fixed error occurring when non-estimable coefficients existed in the component models (problem reported by Annabel Smith)

Version: 1.6.5 (2011-11-28) pdredge example changed to use SOCK cluster type (portability issue).

Version: 1.6.4 (2011-11-26) coefTable: replaces 'tTable' (which still exists, but is deprecated), it accepts an argument 'dispersion'. The dispersion parameter can be also passed from 'model.avg'. pdredge: collects all the warnings and errors which occur during fitting, and stores them. They are printed below the table. The function does more thorough checking of the cluster nodes' environment, for possible missing variables and functions. get.models: "subset" argument can be now a character vector giving model names. (added) demos for model selection with models from 'unmarked' package using 'dredge' and 'pdredge'.

Version: 1.6.3 (2011-11-15) pget.models: new function, parallel version of 'get.models'. (p)get.models: important change of the default behaviour: without the 'subset' argument, ALL models are returned. Also the '...' arguments were not used to update the returned models - this is corrected now.

Version: 1.6.2 (2011-11-13) pdredge: new function, a parallel version of dredge (experimental).

Version: 1.6.1 (2011-11-11) dredge: added generic support for all 'unmarked' models, and fixed some bugs with it. model.names: new function (previously was internal only) CAICF: adds to the collection of rarely used information criteria.

Version: 1.6.0 (2011-11-07) dredge: most important change (although invisible to the user) is in the way the model objects are updated. The new function 'makeArgs' is used to generate the updated call. This results in a very flexible interface to the modelling functions. The default method updates the 'formula' argument, other methods can be used to deal with more problematic calls (the models from package 'unmarked' were a challenge). Apart from the model classes previously supported by 'dredge', support for 'unmarkedFitOccu', 'unmarkedFitColExt' and several other 'unmarkedFit' objects has been added. dredge: has new argument 'extra' to add additional statistics into the model selection table. r.squared: new function to calculate likelihood-ratio based pseudo-R-squared subset method for 'model.selection': accepts a shorthand notation "has(x)" to select rows where a variable is present. (added) support for 'survreg' and 'rq' models (packages 'survival' and 'quantreg')

Version: 1.5.2 (2011-10-21) dredge: fixed a bug with wrong ordering of column names in the returned data.frame (thanks to Staffan Roos for reporting this).

Version: 1.5.1 (2011-10-20) compatibility with R prior to 2.13.0: added workarounds for missing generic function 'nobs'.

Version: 1.5.0 (2011-10-18) model.avg: removed the argument "method", the standard errors and confidence intervals are calculated only for coefficients averaged within a subset of models where they are present (method = "NA"). The "full" coefficients are also returned, but without std. errors. plot.model.selection: new function to visualize the model selection table manuals expanded, examples revised getAllTerms: should work for most of the classes derived from 'unmarkedFit' a load of small improvements, polishings and fixes, generally invisible to the user.

Version: 1.4.4 (2011-10-16) QAIC: no longer uses model deviance, as it gave biased values. ICOMP, Cp: new information criteria model.avg: now performs a smarter check for duplicate models, fixed calculation of importance values model.sel: new alias to 'mod.sel' (eventually will replace 'mod.sel').

Version: 1.4.3 (2011-10-13) Beetle: new data - beetle mortality, with examples from B&A's book dredge: new arguments 'varying', 'm.min' and 'evaluate' manuals modified and corrected tests/gamm.R: tests for 'gamm'/'gamm4' support.

Version: 1.4.2 (2011-10-11) support for 'gamm' and 'gamm4' gamm: added a wrapper for gamm, that allows for updating the object (hopefully needed only temporarily)

Version: 1.4.1 (2011-10-10) A few changes to vignette 'gamm'

Version: 1.4.0 (2011-10-05) importance: new function mod.sel: added method for a 'model.selection' object

Version: 1.3.10 (2011-09-23) Reverted changes made in previous version (package dependencies) to succesfully pass checks.

Version: 1.3.9 (2011-09-17) removed *.lmer *.glmer methods, added dependency on lme4 >= 0.999375-16 tTable.mer: minor change

Version: 1.3.8 (2011-09-17) Restored compatibility with R 2.12.2 getAllTerms: new attribute 'intercept'

Version: 1.3.7 (2011-09-16) handling double intercept (phi(Int) and p(Int)) in 'unmarked'

Version: 1.3.6 (2011-09-12) changes in the documentation and vignette

Version: 1.3.5 (2011-08-05) mod.sel: new function added methods for 'unmarkedFit' objects (package 'unmarked'), to provide (limited) support in 'model.avg'

Version: 1.3.4 (2011-08-03) Added vignette document on model selection with 'gamm' and 'gamm4' (thanks to Graham MacDonald for an incentive)

Version: 1.3.3 (2011-08-01) dredge: small improvements, no longer dependent on 'bitops' Polished the documentation

Version: 1.3.2 (2011-08-01) nobs: updated methods

Version: 1.3.1 (2011-07-31) dredge: bugs fixed

Version: 1.3.0 (2011-07-29) model.avg: now may be used directly with 'model.selection' object summary.averaging: gives now more information

Version: 1.2.4 (2011-07-29) fixed handling of spatial parameters (lambda, rho) in models from package 'spdep' dredge: modified for better performance dredge, get.models: can deal with glms with starting values provided ('start' argument) QAIC: added a note in documentation stating that is uses deviance rather than logLik in the calculation

Version: 1.2.3 (2011-04-07) fixed some REML/ML issues, AICc rewritten (based on the new AIC code in stats)

Version: 1.2.2 (2011-04-06) nobs: corrected methods for mixed models, added documentation page

Version: 1.2.1 (2011-04-05) .GetLogLik - hidden helper function (returns the proper logLik to use)

Version: 1.2.0 (2011-04-04) AICc has new argument REML used with mixed models. Thanks to Benjamin Augustine for suggestion.

Version: 1.1.2 (2011-03-19) removed arguments 'alpha' and 'level' from model.avg. 'confint' is used instead to obtain CI

Version: 1.1.1 (2011-03-18) the printed output of 'averaging' has been made consistent with that of lm and glm: 'print' gives concise information and 'summary' gives more details. confint.averaging: new function. print.summary.averaging: new function.

Version: 1.1.0 (2011-03-10) added z-statistic, p-values (with significance stars) to model.avg output

Version: 1.0.0 (2011-02-10) get.models: fixed - arguments in ... were not passed to updated models predict.averaging: predictions on response scale can be calculated (type=response) new functions (not exported): nobs, coefDf model.avg: now makes use of correct DF from component models in calculation of adjusted SE par.avg: calculates ASE if df are provided, example(QAIC) now demonstrates predictions with confidence intervals

Version: 0.14.2 (2011-02-05) dredge: small changes, slightly faster manual corrected

Version: 0.14.1 (2011-02-04) vcov.averaging, logLik.averaging: new methods, revised formula for unconditional variance manual reworked

Version: 0.14.0 (2011-01-19) TODO: model.avg: added npar to the selection table predict.averaging: now can return averaged SE par.avg: may use revised formula for calculating averaged SE

Version: 0.13.21 (2011-01-16) dredge: handling of update'd lmer models improved QAICc: new function

Version: 0.13.20 (2011-01-16) Namespace exports updated (coxph methods) dredge: uses now QAIC for glms with 'quasi*' family, and if no rank is specified

Version: 0.13.19 (2011-01-15) model.avg: fixed wrong ordering of model names (thanks to Dennis Jonason for noticing this)

Version: 0.13.18 (2010-10-16) Support for coxph (package survival)

Version: 0.13.17 (2010-09-13) model.avg: fixed a bug with models with binary response; dredge: added checking for na.omit'ting in the global model. new tests: varia.R

Version: 0.13.16 (2010-09-07) dredge, model.avg: broken support for several types of models was fixed

Version: 0.13.15 (2010-09-06) support for 'MASS::multinom'

Version: 0.13.14 (2010-09-03) print.model.selection: empty columns are not printed now

Version: 0.13.13 (2010-09-02) dredge: handling of models with no intercept, added "call" attribute update.model.selection: new function

Version: 0.13.12 (2010-08-29) dredge, getAllTerms, dredge.Rd: small modifications

Version: 0.13.11 (2010-08-28) getAllTerms: improved handling of interactions, random terms (lmer) and "offset()" formulaAllowed: now allows for exceptions in margins presence checking (new argument: "except") dredge: now tries to call fitting function directly rather than through update (more effective in case when model not passed as a variable). New arguments: "trace", and "marg.ex" - marginality checking exceptions.

Version: 0.13.10 (2010-08-26) getAllTerms: fixed a bug with with wrong ordering of interaction terms (resulted with a:b or b:a depending on the set of terms). Improved handling of random terms in lme.

Version: 0.13.9 (2010-08-24) dredge: fixed bugs introduced in the previous version

Version: 0.13.8 (2010-08-03) QAIC: examples expanded

Version: 0.13.7 (2010-08-02) dredge: prints errors generated by models as warnings (previously these were skipped silently)

Version: 0.13.6 (2010-08-01) print.model.selection: improved column abbreviation.

Version: 0.13.5 (2010-07-31) dredge: fixed a bug causing interaction coefficients to be stripped from model.selection table, new function 'fixCoefNames' (not exported) handles proper ordering of interaction components in coefficient names.

Version: 0.13.4 (2010-07-21) 'beta.weights' now uses 'tTable' (support for more model types), 'dredge' with beta=TRUE tries out if it can get beta.weights working. Manual: added gls and rlm to the list of supported models

Version: 0.13.3 (2010-07-15) Bug fixed in print.averaging

Version: 0.13.2 (2010-07-14) Attributes for custom rank function in "dredge" and "model.avg" are now evaluated model-wise, if provided as expressions.

Version: 0.13.0-1 (2010-07-12) Fixed compatibility issues with models of class "lme" and "sarlm". "model.avg" does now some checking whether all models were fitted to the same data and have the same response.

Version: 0.12.13 (2010-07-09) 'recalc.weights' argument added to "subset.model.selection". Small bug fixed in dredge.

Version: 0.12.12 (2010-07-08) New in "dredge": Subsetting models a priori according to a formula

Version: 0.12.11 (2010-07-07) Subset/extract methods for 'model.selection' object

Version: 0.12.10 (2010-07-05) Fixed "predict.averaging" for missing 'newdata' handling; Table returned by "dredge" now has factors for terms with more than one level (they were previously shown as '1')

Version: 0.12.9 (2010-07-03) Finally fixed problems with variable exporting and hidden methods

Version: 0.12.7-8 (2010-07-02) Small changes in namespace and documentation

Version: 0.12.6 (2010-06-26) Tidying up the code, small changes.

Version: 0.12.4 (2010-06-21) added: "predict.averaging" and "coef.averaging", 'averaging' object contains additional elements, so that several default methods work with it.

Version: 0.12.2 (2009-06-14) fixed: 'alpha' in "model.avg" was not passed to "par.avg"

Version: 0.12.1 (2009-04-20) fixed: method="NA" in model.avg, par.avg now accepts NA's in 'weight'

Version: 0.12.0 (2009-03-24) Change in version number only. Had to make this jump in numbering, as there was a typo in DESCRIPTION file of an early version, (0.11.2 instead of 0.1.0), and since then R-forge mechanism seemed to ignore actual newer versions, as they had lower number.

Version 0.1.1 (2009-03-22) Modified: dredge: added "fixed" and "m.max" arguments.

Version 0.1.1 (2009-03-17) Fixed: par.avg: Variance estimate corrected (previous estimate was square root of variance). Thanks to Mick Wu. Fixed: dredge: Apparently "expand.grid" cannot handle too many combinations, so now dredge stops with a more explanatory message

Version 0.1.0 dRedging (0.11.2) moved to R-Forge. Since model averaging is now the main focus of the package, its name changed to MuMIn (from MUlti Model INference).