bmm 1.0.0
New features
- you can now specify to save the bmmfit object generated by bmm() to a file with the file argument, similarly to brms::brm() (#190)
- the parameterization of the imm was adapted to accurately reflect the model as implemented by Oberauer et al. (2017)
- prepare package for CRAN submission
Bug fixes
- fix incorrect specification of default priors when only an interaction is specified (#201)
- the random generation function for the mixture3p and imm returned incorrect samples for some rare parameter combinations, this has now been fixed, so that the functions now return correct samples for all parameter combinations.
Deprecated functions and arguments
- BREAKING CHANGE: the arguments for the distribution functions of the mixture2p and mixture3p model have been change to match the snake_case coding scheme. Instead of pMem and pNT these are now p_mem and p_nt. The old names are deprecated and are no longer supported
bmm 0.5.0
New features
- add a summary() method for bmmfit objects (#144)
- add a global option bmm.summary_backend to control the backend used for the summary() method (choices are “bmm” and “brms”)
- function restructure() now allows to apply methods introduced in newer bmm versions to bmmfit objects created by older bmm versions
- you can now specify any model parameter to be a constant by using an equal sign in the bmmformula (#142)
- you can now choose to estimate parameters that are fixed to a constant by default for all models (#145)
- default priors for all models are now specified via the configure_prior() S3 method (#145)
- cmdstanr will be used as the default backend for brms if the user has it installed (#145)
- various updates to the documentation and data sets
Documentation
- two new online articles that introduce the bmmformula syntax and explain how to extract information from bmmodels such as the generated Stan code and Stan data for each model
Deprecated functions and arguments
- BREAKING CHANGE: remove get_model_prior(), get_stancode() and get_standata(). Due to recent changes in brms version 2.21.0, you can now use the brms functions default_prior, stancode and standata directly with bmm models.
- the function fit_model() is deprecated in favor of bmm() and will be removed in a future version (#163)
- the argument setsize for the mixture3p and IMM models is now called set_size for consistency. The old argument name is deprecated and will be removed in a future version (#163)
- the distributions functions for the imm model are renamed from dIMM, pIMM, rIMM and qIMM to dimm, pimm, rimm and qimm (#163)
- the argument parallel for the bmm() function is deprecated and will be removed in a future version. Use cores instead, as for brms::brm() (#163)
- the models IMMfull(), IMMabc() and IMMbsc() are now called via imm(), imm(version = “abc”) or imm(version = “bsc”). The old names are deprecated and will be removed in a future version (#163)
- the sdmSimple() model is now called sdm(). The old name is deprecated and will be removed in a future version (#163)
bmm 0.4.0
New features
- add a check for the sdmSimple model if the data is sorted by predictors. This leads to much faster sampling. The user can control the default behavior with the sort_data argument (#72)
- the mixture3p and IMM models now require that the intercept must be suppressed when set size is used as a predictor (#96).
- add postprocessing methods for sdmSimple to allow the use of pp_check(), conditional_effects and bridgesampling with the model (#30)
- add informed default priors for all models. You can always use the get_model_prior() function to see the default priors for a model
- add a new function set_default_prior for developers, which allows them to more easily set default priors on new models regardless of the user-specified formula
- you can now specify variables for models via regular expressions rather than character vectors (#102)
- you can now view and set all bmm global options via bmm_options(). See ?bmm_options for more information
- add a start-up message upon loading the package
Bug fixes
- fix a bug in the mixture3p and IMM models which caused an error when intercept was not suppressed and set size was used as predictor
- update() now works properly with bmmfit objects (#95)
- fix a bug in the sort_data check which caused an error when using grouped covariance structure in random effects across different parameters
bmm 0.3.0
New features
- BREAKING CHANGE: The fit_model function now requires a bmmformula to be passed. The syntax of the bmmformula or its short form bmf is equal to specifying a brmsformula. However, as of this version the bmmformula only specifies how parameters of a bmmodel change across experimental conditions or continuous predictors. The response variables that the model is fit to now have to be specified when the model is defined using model = bmmodel(). (#79)
- BREAKING CHANGE: The non_target and spaPos variables for the mixture3p and IMM models were relabeled to nt_features and nt_distances for consistency. This is also to communicate that distance is not limited to spatial distance but distances on any feature dimensions of the retrieval cues. Currently, still only a single generalization gradient for the cue features is possible.
- This release includes reference fits for all implemented models to ensure that future changes to the package do not compromise the included models and change the results that their implementations produce.
- The check_formula methods have been adapted to match the new bmmformula syntax. It now evaluates if formulas have been specified using the bmmformula function, if formulas for all parameters of a bmmodel have been specified and warns the user that only a fixed intercept will be estimated if no formula for one of the parameters was provided. Additionally, check_formula throws an error should formulas be provided that do not match a parameter of the called bmmodel unless they are part of a non-linear transformation.
- You can now specify formulas for internally fixed parameters such as mu in all visual working memory models. This allows you to predict if there is response bias in the data. If a formula is not provided for mu, the model will assume that the mean of the response distribution is fixed to zero.
- there is now an option bmm.silent that allows to suppress messages
- the baseline activation b was removed from the IMM models, as this is internally fixed to zero for scaling and as of now cannot be predicted by independent variables because the model would be unidentifiable.
- the arguments used to fit the bmmodel are now accessible in the bmmfit object via the
fit$bmm$fit_argslist. - add class(‘bmmfit’) to the object returned from fit_model() allowing for more flexible postprocessing of the underlying brmsfit object. The object is now of class(‘bmmfit’, ‘brmsfit’)
- changes to column names of datasets zhang_luck_2008 and oberauer_lin_2017 to make them more consistent
Bug Fixes
- an error with the treatment of distances in the IMMfull and the IMMbsc has been corrected. This versions ensures that only positive distances can be passed to any of the two models.
- removed a warning regarding the scaling of the distances in the IMMfull and the IMMbsc that was specific only for circular distances.
bmm 0.2.2
Bug Fixes
- fixed a bug where passing a character vector or negative values to set_size argument of visual working memory models caused an error or incorrect behavior (#97)
bmm 0.2.0
New features
- New model available - The Signal Discrimination Model by Oberauer (2023) for visual working memory continuous reproduction tasks. See ?sdmSimple. The current version does not take into account non-target activation
- Add ability to extract information about the default priors in bmm models with get_model_prior() (#53)
- Add ability to generate stan code and stan data for each model with get_model_stancode() and get_model_standata() (#81)
- BREAKING CHANGE: Add distribution functions for likelihood (e.g. dIMM()) and random variate generation rIMM()) for all models in the package. Remove deprecated gen_3p_data() and gen_imm_data() functions (#69)
- Two new data sets are available: zhang_luck_2008 and oberauer_lin_2017 (#22)
Documentation
- Website for the development version of the package is now available at https://venpopov.github.io/bmm/dev/ (#18)
- Add articles for each model to the website at https://venpopov.github.io/bmm/dev/articles/
- Add a detailed developer’s guide to the website at https://venpopov.github.io/bmm/dev/dev-notes (#21)
- Improve README with more detailed information about the package’s goals and its models (#21)
Other changes
- Save bmm package version in the brmsfit object for reproducibility - e.g.
fit$version$bmm(#88)
bmm 0.1.1
New features
- BREAKING CHANGE: Improve user interface to fit_model() ensures package stability and future development. Model specific arguments are now passed to the model functions as named arguments (e.g. mixture3p(non_targets, setsize)). This allows for a more flexible and intuitive way to specify model arguments. Passing model specific arguments directly to the fit_model() function is now deprecated (#43).
- Add information about each model such as domain, task, name, version, citation, requirements and parameters (#42)
- Add ability to generate a template file for adding new models to the package with use_model_template() (for developers) (#39)
bmm 0.1.0
A major restructuring of the package to support stable and generalizable development of future models (#41).
New Features
- Refactor the fit_model() function to be generic and independent of the model being fit (#20)
- Transform models to be S3 objects. (#41).
- View currently supported models with new function supported_models(). Currently supported models are: mixture2p(), mixture3p(), IMMabc(), IMMbsc(), IMMfull()
- Add S3 methods for checking the data, formula, model and priors (#41)
- Add distribution functions for the Signal Discrimination Model. See ?SDM for usage (#27)
- Add softmax and softmaxinv functions
Bug Fixes
- Change default prior on log(kappa) to Normal(2,1) for the mixture3p() model (#15)
Other changes
- BREAKING CHANGE: deprecate model_type argument in fit_model(). Models must now be specified with S3 functions passed to argument model rather than model names as strings passed to argument model_type (#41)
- Add extensive unit testing