measr 2.0.0

Breaking changes

• The S3 parts of measr have been converted to {S7}.

• Some of measr’s functionality has been decoupled into other packages to allow for quicker and easier updates. The generation of Stan code and data lists have been moved to {dcmstan}. Similarly, the example data sets have been moved to {dcmdata} to facilitate the use of the data across other packages. As part of the decoupling, measr_dcm() has been deprecated in favor of dcmstan::dcm_specify() and dcm_estimate().

  # old
  my_model <- measr_dcm(
    data = ecpe_data,
    qmatrix = ecpe_qmatrix,
    resp_id = "resp_id",
    item_id = "item_id",
    type = "lcdm",
    attribute_structure = "unconstrained",
    method = "mcmc"
  )
  
  # new
  library(dcmdata)
  
  my_spec <- dcm_specify(
    qmatrix = ecpe_qmatrix,
    identifier = "item_id",
    measurement_model = lcdm(),
    structural_model = unconstrained()
  )
  my_model <- dcm_estimate(
    dcm_spec = my_spec,
    data = ecpe_data,
    identifier = "resp_id",
    method = "mcmc"
  )

• predict() has been deprecated in favor of score(). The functionality is the same, but

New features

• aic() and bic() have been added for estimating relative model fit for models estimated with method = "optim" (@JeffreyCHoover, #54).

• bayes_factor() has been added for comparing models using Bayes factors. This is only available for models estimated with backend = "rstan" (@JeffreyCHoover, #67).

• Item and attribute discrimination measures can now be calculated with cdi() (@auburnhimenez34, #63).

• The specified Q-matrix can now be evaluated and compared to other empirical Q-matrix specifications using qmatrix_validation() (@JeffreyCHoover, #65).

• In reliability(), users can now calculate the classification accuracy and consistency for different probability classification threshold by specifying a threshold (#45).

• New estimation methods, variational() and pathfinder(), have been added to support estimation via Stan’s variational algorithm for approximate posterior sampling and the pathfinder variational inference algorithm, respectively. Pathfinder is only available when the model is estimated with {cmdstanr} (#72).

• Local item dependence can now be estimated with yens_q3() (@JeffreyCHoover, #62).

Minor improvements and fixes

• Documentation has been updated to ensure examples use Air formatting to improve accessibility (#68).

• measr_extract() has been updated to no longer require adding elements to a model object before extracting (#73).

measr 1.0.0

New documentation

• A new article on model evaluation has been added to the project website (https://measr.r-dcm.org).

• The model estimation article has been updated to use the same (simulated) data set as the model evaluation article.

• More detailed installation instructions have been added to the getting started vignette (#23).

• A case study demonstrating a full DCM-based analysis using data from the ECPE (?ecpe_data) has been added to the project website.

Minor improvements and fixes

• Fixed bug in the LCDM specification of constraints for level-3 and above interaction terms.

• Functions for evaluating estimated models (e.g., fit_ppmc(), reliability()) no longer recalculate indices if they have previously been saved to the model object. This behavior can be overwritten with force = TRUE.

• Updated Stan syntax to be compatible with the new array syntax (@andrjohns, #36)

• get_parameters() now preserves item identifiers by default. Items can be renamed with numbers (e.g., 1, 2, 3, …) by setting rename_item = TRUE.

• measr now reexports functions from posterior for conducting mathematical operations on posterior::rvar() objects.

• Respondent estimates are now returned as posterior::rvar() objects when not summarized.

measr 0.3.1

• Added a NEWS.md file to track changes to the package.

New features

• Support for additional model specifications has been added (#10):
  • The compensatory reparameterized unified model (C-RUM) can now be estimated by defining type = "crum" in the measr_dcm() function.
  • Users can now drop higher order interactions from the loglinear cognitive diagnostic model (LCDM). A new argument for measr_dcm(), max_interaction, defines the highest order interactions to estimate. For example, max_interaction = 2 will estimate only intercepts, main effects, and two-way interactions.
  • A new argument to measr_dcm(), attribute_structure allows users to specified either “unconstrained” relationships between attributes or “independent” attributes.
• Updated prior specifications:
  • Users can now specify a prior distribution for the structural parameters that govern the base rates of class membership (#2).
  • Safeguards were added to warn users when a specified prior is not defined for the chosen DCM sub-type. For example, an error is generated if a prior is defined for a slipping parameter, but the LCDM was chosen as the type of model to be estimated (#1).

Minor improvements and fixes

• Fixed bug with backend = "rstan" where warmup iterations could be more than the total iterations requested by the user if warmup iterations were not also specified (#6).

• Additional specifications were added to measr_extract() for extracting results from an estimated model.