psychtm 2021.1.0

• Initial release for CRAN.

Major changes

• Requires R packages:
  • Requires coda v0.4 or higher.
  • Requires Rcpp v0.11.0 or higher.
  • Requires tibble v2.1.3 or higher.
• Added a vignette to illustrate package use: closes (#47).
• Added example data set teacher_rate for demonstration in vignette.
• New function prep_docs() to prepare documents in a data frame for modeling with gibbs_sldax().
• New function post_regression() to summarize regression relationships for objects of class Mlr, Logistic, or Sldax: closes (#7).

Minor changes

• Added a NEWS.md file to track changes to the package.
• Suggests R packages:
  • Suggests spelling.
  • Suggests knitr v1.22 or higher.
  • Suggests lda
  • Suggests testthat v3.0.2 or higher.
  • Suggests rmarkdown.
• Uses roxygen2 v7.1.2 with Markdown support for documentation.
• Expanded and improved documentation.
• Remove user-facing documentation in man/ for internal functions.
• Language for package now listed as en-US.
• S4 class definitions moved to a single file.
• S4 generic functions moved to a single file.
• S4 methods for Sldax objects moved to a separate file.
• Deprecated gg_coef(); this function will be removed in a future release.
• Use of stop() and warning() in the event of errors or warnings instead of print() and cat().
• Use of Rcpp_cout and Rcpp_cerr for printing messages and errors from compiled functions.
• Improved README.
• Added CITATION information.

Bug fixes

• Fixed namespace issues for importing and exporting packages, methods, and functions.
• Corrected bug when fitting an SLDAX model with only a single predictor where variable names were accidentally omitted.
• Fixed bug where get_zbar.Sldax() was missing a return statement.
• Makevars.win no longer asks for OpenMP as OpenMP is not currently used.

psychtm 2020.9-alpha

Non-breaking Changes

• Fixed bug when correcting label switching for an LDA model.

psychtm 2020.8-alpha

Non-breaking Changes

• Fixed bug when correcting label switching when manifest predictors for an SLDAX model were included where the manifest predictors’ regression coefficients were not stored in the model object.

psychtm 2020.6-alpha

Non-breaking Changes

• Fixes to Travis CI build configuration file.
• Updated README.

psychtm 2020.5-alpha

Non-breaking Changes

• Minor documentation fix.

psychtm 2020.4-alpha

New Features

• Coherence (Mimno, Wallach, Talley, Leenders, & McCallum, 2011) and exclusivity (Roberts, Stewart, & Airoldi, 2013) metrics of topic model fit can now be computed:
  • get_coherence() computes the coherence score for each topic,
  • get_exclusivity() computes the exclusivity score for each topic.
• Label switching correction using the Stephens (2000) algorithm is now supported for SLDAX and other topic models: closes (#42).
  • gibb_sldax() gains the argument correct_ls (TRUE / FALSE),
  • A logical flag indicating whether a label switching correction was applied is now stored in @extra$corrected_label_switching for objects of class Model.
• Added options to sample the topic proportions and topic-word probabilities or just obtain a point estimate to gibbs_sldax():
  • Argument sample_theta (default = TRUE) for the topic proportions,
  • Argument sample_beta (default = TRUE) for the topic-word probabilities.
• Added option to store the chain of sampled topic assignments to gibbs_sldax().
  • Argument return_assignments (default = FALSE).
  • Keeping this FALSE is dramatically more memory efficient: closes (#37).
• Added a check to ensure that the number of observations supplied to docs and data arguments of gibbs_sldax() are equal since missing data in either argument is currently not supported: closes (#36).
• Extractor functions are now available to safely obtain the contents of object slots: closes (#1).
  • For example, to obtain the posterior samples of regression coefficients from a fitted SLDAX model called my_fit, use eta(my_fit).

Breaking Changes

• Fixed WAIC calculations: closes (#11).
• Created hierarchical S4 classes that inherit slots as needed from parents (may not be a breaking change for many users):
  • Model: a generic model class,
  • Mlr: a linear regression class which inherits from Model,
  • Logistic: a logistic regression class which inherits from Model,
  • Sldax: a topic model class which inherits from Mlr and Logistic.

Non-breaking Changes

• Imports label.switching package.
• Changed default prior hyperparameters in gibb_sldax() for the Dirichlet priors on the topic proportions and on the topic-word probabilities to 1.0 (flat over the simplex).
• Minor fix in calculation for when to message user that burn-in period is finished if verbose = TRUE specified to gibbs_sldax().

psychtm v2020.3-alpha

New Features

• Fixed error in gibbs_*() functions where interrupting the process (e.g., hitting the Esc key while running) and then trying to run these functions again would fail with the error ERROR: there is already an InterruptableProgressMonitor instance defined.

Breaking Changes

• Requires RcppProgress v0.4.2.

Non-breaking Changes

• None.

psychtm v2020.2-alpha

New Features

• Fix initialization error in gibbs_logistic() where regression coefficients could start with infinite values and the algorithm could not recover: closes (#39).

Breaking Changes

• None.

Non-breaking Changes

• See above in New Features.

psychtm 2020.1-alpha

New Features

• Better error handling for gibbs_*() functions.
• Increased maximum number of iterations attempted from 1,000 to 10,000 when trying to sample regression coefficients in gibbs_sldax() with constrain_eta = TRUE and now warns user if exceeded.

Breaking Changes

• None.

Non-breaking Changes

• Import Rcpp, RcppArmadillo, and RcppProgress packages.
• Split up C++ codebase into multiple files for easier maintenance going forward and faster compilation if using ccache.

psychtm 2019.7-alpha

New Features

• Added unit tests (still a work in progress to cover entire package).
• Documentation fixes for a number of functions: closes (#33).
• Faster computation for get_zbar().
• Better argument checks for some functions.

Breaking Changes

• Intermediate C++ functions are no longer exported: .est_betak(), .est_thetad(), .count_topic_word(), .gibbs_mlr_cpp(), .gibbs_logistic_cpp(), .gibbs_sldax_cpp(), rmvnorm_cpp(), pwaic_d(), waic_d().
• get_toptopics() no longer requires a Sldax model object and instead needs a topic-proportions matrix for all documents (‘theta’): closes (#32).
• Removed obsolete classes (LDA, Slda, Sldalogit) and reorganized Sldax class.

Non-breaking Changes

• Changes to est_beta()/est_theta(): remove unwanted dim name in returned matrix; handle error in case where chain was shorter than 10 iterations.
• gg_coef() no longer automatically prints a plot if the result of gg_coef() is assigned to an object.
• Vectorized computation in term_score() for 2-3x speed improvement, especially for large matrices.
• Import is() from methods package.