2022-10-23:
  - Refactored MQL/PQL algorithm: Eliminated redundant code and adapted it to
    both 'nlm' and 'nlminb'

2022-10-16:
  - Fixed bug in MQL/PQL-objective function that led to false non-convergence and
    bias in variance parameter estimates

2022-10-12:
  - Support for starting values in random effects models
  - Support for restriction on random effects variances in multinomial baseline
    logit models

2022-10-09:
  - Improve handling of boundary values and singular information matrices

2022-10-07:
  - Remove spurious messages about missing starting values

2022-05-21:
  - Add checks of 'control=' argument of 'mclogit()' and 'mblogit()'.

2022-04-13:
  - Fixed bug in 'blockMatrix' and make it check for argument validity

2022-04-11:
  - Hotfix of prediction method

2022-04-10:
  - Fix handling of singular initial covariance matrices in PQLMQL_innerFit
  - Issue a warning if models with random effects are compared using anova
  - Fix predict methods for mmclogit models
  - Handle DOIs in documentation as required by new guidelines

2022-01-16:
  - Fix prediction with complicated terms in the model
  - Add some more demos

2021-08-13:
  - predict.mmclogit: create W-Matrix only when really needed

2021-07-13:
  - Include variance parameters in the computation of degrees of freedom

2021-06-03:
  - Be less zealous about group-level covariates constant in some choice sets.

2021-05-30:
  - Added support for vertical-bar syntax for responses of conditional
    logit models.

2021-05-27:
  - Added support for non-nested random effects.

2021-05-25:
  - Fixed serious bug in the handling of multilevel random effects models.
  - Detect some misspecified models with too many groups.

2021-04-17:
  - Merged pull request by Russel V. Lenth that adds support for "emmeans".

2021-04-04:
  - Apply patch suggested by Ilya Yalchyk to improve formula argument
    of 'mclogit()' and 'mblogit()'.

2021-03-19:
  - Last fixes for CRAN

2021-03-18:
  - Improved support 'mtable()' for multinomial logit models with random effects.

2021-02-21:
  - Fixed predictions from models with scaled independent variables etc.
  - 'summary()' now reports the number of groups per random effects level.      

2021-01-28:
  - Another prediction fix. Do not refer to weights that are not needed.

2021-01-10:
  - Fixed prediction method also for mmclogit objects

2020-12-23:
  - Refactored computations
  - Fixed predictions from random-effects models where group indices are not 
    a sequence of integers starting at 1.

2020-11-03:
  - Correct URLs in DESCRIPTION

2020-09-09:
  - Fix reference to weights in 'predict()' methods

2020-08-06:
  - Let 'mclogit'/'mblogit' handle empty responses (i.e. where counts sum to
    zero) correclty.
  - Make 'mclogit' complain about non-numeric responses  

2020-07-17:
  - Documented prediction methods.
  - Improved flexibility of prediction methods. 

2020-07-16:
  - Implemented reasonable 'predict' method for mmblogit and mmclogit objects.

2020-07-15:
  - Bugfix: Make 'update' work with missing 'dispersion=' argument.
  - Bugfix: Make 'vcov' work for objects without 'phi' component.
  - Add 'vcov' method for 'mmclogit' objects.

2020-06-27:
  - Documented 'simulate()' methods.

2020-06-11:
  - Implemented (approximate) REML estimator.

2020-06-07:
  - Added a 'simulate()' method for "mblogit" and "mclogit" models.

2020-05-24:
  - Adapt the package NAMESPACE file to explicitly export S3 methods as methods,
    even if they are export as functions, as newly required by R 4.0.

2020-05-23:
  - Added documentation of (over-)dispersion parameter estimation, rename
    'overdispersion=' arguments into 'dispersion=' arguments.

2020-05-22:
  - Added support for estimation of (over-)dispersion parameters

2020-05-21:
  - Implemented MQL technique as an alternative to PQL estimation

2020-05-19:
  - Improve handling of numerical difficulties

2020-05-11:
  - Use a Cholesky-factor parameterisation to make sure that covariance matrices
    are positive (semi-)definite

2020-03-30:
  - Refactored the algorithm for fitting mixed-effects models

2020-01-09:
  - Document getSummary.mmclogit, getSummary.mmblogit
  - Make 'mblogit' handle matrix responses with zero-sum rows
  - Renamed 'residual.df' to 'df.residual' in results object of 'mclogit.fit' et
    al. (Fixes issue #4)

2019-10-23:
  - Merge pull request #3 from skyborla/fix-mblogit
    Fix mblogit for responses with only two levels
  - Export getSummary.mmclogit, getSummary.mmblogit

2019-04-20:
  - Let 'mmblogit' models inherit from class 'mblogit'

2019-02-04:
  - Merged pull request #2 from pmcharrison/ftt-fix:
    Fixed typo (ftt -> fit)

2018-09-26:
  - Fixed matrix column selection in predict.mclogit if there is only one
    predictor (also PR from skyborla)

2018-04-25:
  - Improved handling of with misspecified random effect structures.
  - Added documentation about new handling of misspecified models.

2017-10-25:
  - Fixed handling of dropped predictors in `predict.mclogit`.

2017-01-26:
  - Fixed some bugs in predict models for `mclogit` objects.
  - Made sure that dummy coding is used for response factors even if they are ordinal

2017-01-07:
  - Implemented random slopes for baseline logit models.

2017-01-05:
  - Implemented random slopes for conditional logit models.

2016-09-01:
  - Fixed `mclogit.fit()` and `predict.mclogit()` to work better without covariates.

2016-02-07:
  - Explicitely import package "methods"

2016-01-17:
  - Import `as` from package "methods".
  - Make sure `nobs` is defined in `mclogit.fit.rePQL`.

2016-01-16:
  - Updated `DESCRIPTION` file: Maintainer email address changed and no "This package"
    at start of package discriptions.

2015-10-08:
  - Fix display of number of observations
  - Drop redundant coefficients

2015-08-01:
  - Added row and column names to estimator result of `vcov()`
  - Make sure that scripts run with "mclogit" loaded by `devtools::load_all()`

2015-07-15:
  - mclogit, mclogit.fit: Added support for starting values.

2015-07-03:
  - predict.mblogit: 'contrasts.arg' not 'contast.arg' ...
  - predict-methods now should handle NAs in newdata arguments better.

2015-06-17:
  - Corrected handling of weights and standard errors of prediction.

2015-06-15:
  - 'getSummary' methods now return "contrasts" and "xlevels" components.
  - Fixed prediction method for 'mclogit' results.
  - Added 'fitted' and 'predict' methods for 'mblogit' results.

2015-01-25:
  - Added support for multinomial baseline logit models in form of 'mblogit' as a frontend to 'mclogit.fit'

2015-01-23:
  - Added URLs to DESCRIPTION file

2015-01-21:
  - Added `nobs` and `extractAIC` methods for `mclogit` objects, so that `drop1.default` should work with these.

2015-01-19:
  - Added call to result of `getSummary.mclogit`.

2015-01-18:
  - Cleanup of NAMESPACE file; added aliases to methods for `mclogit` objects so that users can see that they are present.
  - Export `mclogit.fit`, and `mclogit.fit.rePQL` to enable use by other packages.

2014-10-13:
  Simplified some namespace dependencies. Eliminated useless pseudo-R-squared statistics
  from getSummary.mclogit

2014-08-23:
  Added 'anova' methods

2014-03-10:
  Refactored code -- algorithms should be more transparent and robust now (hopefully!).
  mclogit without and with random effects can handle missing values now.
  Fixed predict method -- use of napredict; handles single indep-variable situation now.
  Fixed embarassing typo -- prior weights do work now (again?).
  Included AIC and BIC methods contributed by Nic Elliot <nic_elliot@yahoo.co.uk>
