major update, mlogit now depends on dfidx, mlogit.data, mFormula and index are deprecated
the name of the coefficients is changed, i.e. air:income is now income:air
a new package version which coincides with the Journal of Statistical Software article
some bugs fixed in the vignettes (citet/citep replaced by markdown-like commands), the figure is added in the mixed logit vignette
Rout.save files are added in the test directory
effects.mlogit now returns a matrix
the R files for the vignettes are added (they were note while building the package on RForge)
thank’s to Mallick Hossain, a bug is fixed in the logsum function
the Cracker, Catsup and Car data set are back in mlogit since AER, flexmix and mlogitBMA run examples based on them.
the alt vector in the index is now carrefully checked in case of alternative subseting or reference level change.
the main vignette is improved, writen in markdown and now and split by sections
the Exercises vignette is splited and is now writen in markdown
importantly, the Cholesky matrix is now coerced to a vector by rows and note by columns.
the mlogit function was checked and improved.
implementation of the computation of the standard deviations of the covariance matrix of the random parameters, using the delta method.
some data sets are removed
zbu and zbt distributions are added : these are one-parameter distributions for which the lower bond is 0,
a logsum function is provided to compute the log-sum or the inclusive utility of a random utility model,
group-hetheroscedastic model can be estimated by setting the relevant covariates in the 4th part of the formula,
the linear predictor is now returned by mlogit,
correlation can still be a boolean, but can also be a character vector if one wants that a subset of the random parameters being correlated.
the RiskyTransport data set (used in the vignette to illustrate the estimation of the mixed logit model
the NOx data set (used in the vignette to illustrate the estimation of the multinomial and group-heteroscedastic logit model),
the JapaneseFDI data set (used in the vignette to illustrate the estimation of the nested logit model)
A new vignette called mlogit2 is added ; this is the draft version of an article submitted to the Journal of Statistical Software ; it is less exhaustive, but better writen thant the original mlogit vignette.
the id series (one observation per choice situation) was badly constructed, it is now fixed
the levels of the choice variable are now equalized to the those of the alt variable, allowing the case were some alternatives are never chosen
mlogit is now able to estimate models with singular matrix of covariates. At the end of model.matrix.mformula, the linear dependent columns of X are removed
there was a bug in the triangular distribution which is now fixed
bug in the effects method fixed
the list of primes used to generate halton sequences was too short, its length has been increased
halton sequences where used to estimate mixed logit even for the default value of halton (NULL), this has been fixed
the contribution of each observation to the gradient is not returned as the ‘gradient’ element of mlogit objects
the distributions are now checked for rpar and an error is returned in case of unknown distribution
ranked-order models can be now estimated ; a new argument called ‘ranked’ is introduced in mlogit.data which performs the relevant transformation of the data.frame. The estimated model is then a standard multinomial logit model
multinomial probit model is now estimated by setting the new probit arguments to TRUE
for the mixed logit model, different draws are now used for each observation
a predict method is now available for mlogit objects
a coef method is added which removes the fixed argument
constPar can now be a named numeric vector. In this case, default starting values are changed according to constPar
the vcov method for mlogit objects is greatly enhanced.
mlogit objects now have two elements which indicate the fitted probabilities : fitted is the estimated probability for the outcome and probabilities contains the fitted probabilities for all the alternatives
mentions to ‘alt’ in the names of the effects is canceled ; moreover, the intercepts are now called altname:(‘intercept’)
a ‘choice’ attribute is added to mlogit.data objects
an effects method is provided, which computes the marginal effects of a covariate
all the models could normally be estimated on unbalanced data
the three tests are added, i.e. a new scoretest function and specific methods for waldtest and lrtest from the lmtest package
the model.matrix method for mlogit objects is now exported
mFormula modified so that models can be updated
likelihood has been rewriten for the heteroscedastic logit model, the computation is now much faster
nested logit models with overlapping nests are now supported; nests = “pcl” enables the estimation of the pair combinatorial logit model
the norm argument is added to rpar
the logLik argument is now of class logLik
mlogit.data is modified so that an id argument can be used with data in long shape
the argument of mlogit.data used to define longitudinal data is now called id.var
mlogit.lnls is corrected so that the estimation of multinomial models can handle unbalanced data (pb with Reduce)
the three tests are temporary removed
a third part of the formula is added : it concerns alternative specific variables with alternative specific coefficients
improved presentation for the Fishing dataset.
a bug (forgotten drop = FALSE) corrected in model.matrix.mFormula
Electricity and ModeCanada datasets are added
if the choice variable is not an ordered factor, use as.factor() instead of class() <- “factor”
cov.mlogit, cor.mlogit, rpar , med, rg, stdev, mean functions are added to extract and analyse random coefficients.
a panel argument is added to mlogit so that mixed models with repeated observation can be estimated using panel methods or not
a problem with the weights argument is fixed
the estimation of nested logit models with a unique elasticity is now possible using un.nest.el = TRUE
the estimation of nested logit models can now be done with or without normalization depending on the value of the argument unscaled
mlogit didn’t work when the dependent variable was an ordered factor in a “wide-shaped” data.frame.
the reflevel argument didn’t work any more in version 0.1-3.
major change, most of the package has been rewriten
it is now possible to estimate heteroscedastic, nested and mixed effects logit model
the package doesn’t depend any more on maxLik but a specific optimization function is provided for efficiency reason
robust inference is provided with meat and estfunc methods defined for mlogit models.
subset argument is added to mlogit so that the model may be estimated on a subset of alternatives.
reflevel argument is added to mlogit which defines the base alternative.
hmftest implements the Hausman McFadden test for the IIA hypothesis.
mlogit.data function has been rewriten. It now use the reshape function.
logitform class is provided to describe a logit model: update, model.matrix and model.frame methods are available.