$call in IRT functions raschmodel(), rsmodel(), pcmodel(), nplmodel(), gpcmodel() (suggested by Rainer W. Alexandrowicz).Fixed calculation of person parameter covariance matrix estimate (and thus confidence intervals) in personpar() methods for Rasch models (reported by Rainer W. Alexandrowicz).
Fixed bug in threshpar() method of pcmodel objects fitted to binary items reported by Rainer W. Alexandrowicz).
Fixed bug in pcmodel() which failed when the reference category was a null category (without observations). Now the first category with observations is chosen as the reference category (reported by Rainer W. Alexandrowicz).
Improved raschmodel(), rsmodel(), and pcmodel() so that an NA variance-covariance matrix is returned (with a warning) if the Hessian cannot be inverted instead of stopping with a technical error message (suggested by Rainer W. Alexandrowicz).
Fixed names of "mscale" attribute in itemresp variable in MathExam14W data, now consistent with column names.
Bug fix in discrpar() method for nplmodel(..., type = "1PL") so that the variance covariance matrix is computed correctly.
Documentation enhancements in the mathematical notation (suggested by Kurt Hornik) and in the description of the 1PL/Rasch model in nplmodel() (suggested by Feng Lin).
The function (and resulting class) for fitting n-PL type parametric logistic IRT models has been renamed from plmodel() to nplmodel(). For now, a copy called plmodel() is also preserved (but this also returns an object of class "nplmodel").
Fix logicals of length > 1 in some if() constructs.
Accept .replicates for user-provided model specifications in mptmodel().
Changed default constant anchor selection method in anchor() and anchortest() from MPT to Gini based on Strobl et al. (2021, Applied Psychological Measurement). The Gini-based anchor selection is simpler because it is based on single anchors while at the same time performing very well under DIF.
The print() method for anchortest() now just displays the anchor item(s) and the final DIF tests while the summary() method displays the full information including anchored item parameters etc. (rather than vice versa).
The print() and summary() methods for anchor() and anchortest() now display item labels (rather than indexes) for the selected anchor item(s) and for the full vector of criterion values.
Added the "proCNI" multinomial processing tree specification in mptspec() for mptmodel(). This provides the CNI model of moral dilemma judgment for proscriptive norms.
vignette("toolbox-simulation", package = "psychotools") on how to conduct simulation studies investigating the performance of score-based tests of measurement invariance of IRT models. Accompanies the PsyArXiv preprint mentioned below.New anchor() selection strategy using inequality-based alignment, either based on the Gini index or the component loss function (CLF). Thus, also directly available in anchortest(). Improved print() and plot() methods.
Changed method to invert Hessian for raschmodel(), rsmodel(), and pcmodel() from qr.solve() to chol2inv(chol()).
Added two demos on how to conduct simulation studies for score-based test of measurement invariance.
All IRT models now have a function to simulate IRT data, see rrm(), rrsm(), rpcm(), rpl(), and rgpcm().
An accompanying new PsyArXiv Preprint “An R Toolbox for Score-Based Measurement Invariance Tests in IRT Models” by Lennart Schneider, Carolin Strobl, Achim Zeileis, and Rudolf Debelak is available at https://doi.org/10.31234/osf.io/r9w34
Infrastructure for IRT modeling in the unified psychotools framework is extended by marginal maximum likelihood (MML) estimation of generalized partial credit models and parametric logistic models, respectively. The corresponding fitting functions (see below for details) call mirt() or multipleGroup() from the mirt package but return objects for which all standard extractor methods (item parameters, person parameters, etc.) and visualization methods (item response curves, parameter profiles, person-item maps, etc.) are available.
The new gpcmodel() function interfaces mirt (see above) and fits (generalized) partial credit models (GPCMs) by MML.
The new plmodel() function interfaces mirt (see above) and fits various parametric IRT logistic models using MML: 1PL (Rasch), 2PL, 3PL, 3PLu, and 4PL.
New functions and eponymous classes guesspar(), and upperpar() to extract/represent so-called guessing parameters and upper asymptote parameters of IRT models.
personpar() now distinguishes between parameters of the assumed person ability distribution (personwise = FALSE) and the individual person parameters for each person/subject in the underlying data set (personwise = TRUE). In the CML case, the latter simply computes the raw score for each person and then extracts the corresponding person parameter. In the MML case, this necessitates (numerically) integrating out the individual person parameters (also known as factor scores or latent trait estimates) based on the underlying normal distribution.
Added new data set ConspiracistBeliefs2016 from the Open Source Psychometrics Project (2016).
Added new simulated data set Sim3PL for fitting dichotomous IRT models, especially the 3PL and 3PLu.
Conditionally register all estfun() and bread() S3 methods for model objects, provided that the sandwich package is attached.
Added native routine registration for esf.c.
Use R version of elementary_symmetric_functions() by default on Win/i386 due to small numeric differences on that platform.
The estfun() method for btmodel objects always computed the scores with the last object for the reference category - even if a different ref= was specified in the model. (Thanks to Heather Turner for pointing out the problem.)
The itempar() method for btmodel objects miscomputed the variance covariance matrix (unless the first object was used as the ref when estimating the model). (Thanks to Heather Turner for pointing out the problem.)
Added new data set PairClustering from Klauer (2006).
Fixed replication code in example of StereotypeThreat (reported by Ed Merkle).
Basil Abou El-Komboz changed his name to Basil Komboz.
Properly imported grDevices and utils in NAMESPACE.
Added new item response data set MathExam14W with esponses of 729 students to 13 items in a written exam of introductory mathematics along with several covariates.
New function mptmodel() and corresponding extractor functions for fitting multinomial processing tree (MPT) models. These functions are somewhat experimental, and their user interface might change in future releases.
Bug fix in itempar() method for raschmodel objects if alias = FALSE. In the previous version the methods had an erroneous trailing NA.
Improved item names labeling in plot() method for itemresp objects to conform with regionplot() function for IRT models.
mscale<-() method for itemresp has been improved so that categories can be easily collapsed (e.g., dichotomized).
Infrastructure for IRT modeling in psychotools is greatly enhanced. Therefore the main modeling functions are now called raschmodel() for Rasch models, rsmodel() for rating scale models, pcmodel() for partial credit models, and btmodel() for Bradley-Terry models. The old *.fit() functions from previous versions of the package still exist but now internally call the new *model() functions. Also, the classes returned have the same names as the *model functions.
A unified visualization framework for fitted IRT models has been added: For all types of models (Rasch, RSM, PCM) one can visualize profiles of the item parameters, regions for the most likely response, item or category characteristic curves, item information, and person-item plots. All of these rely on the unified framework for extracting parameters and predictions (see below).
New functions and eponymous classes itempar(), threshpar(), and discrpar() to extract/represent item, threshold, and discrimination parameters of item response models. Methods for the IRT models (Rasch, RSM, PCM) are provided. In addition, several methods for standard generic functions (print(), coef(), vcov()) are available.
The worth() generic now internally calls the methods for itempar().
Estimation of person parameters for a given item response model is now available via the generic function personpar(). Specific methods for Rasch, rating scale and partial credit models allow the estimatation of person parameters via joint maximum likelihood estimation. Methods for standard generic functions (print(), coef(), vcov()) are available for the resulting objects of class personpar.
predict() methods for Rasch, rating scale and partial credit models have been added. For a given fitted model object, these can be used to predict various types of response probabilities or actual reponses.
New functions anchor() and anchortest() provide a variety of anchor methods for the detection of uniform differential item functioning (DIF) between two pre-specified groups in the Rasch model. To test for DIF, the itemwise Wald test is implemented.
itemresp() is the class constructor for responses of n subjects to k items which can be polytomous and have different measurement scales. A wide range of methods to standard generics is provided as well as to generics created for the paircomp class. Thus, features can be easily extracted/replaced, summaries/visualizations can be produced, subsetting/merging/etc. is facilitated.
The handling of argument ref when producing a region plot (previously called effect plot) was changed. Whereas in the previous implementation, the restriction specified in this argument was applied to the cumulative absolute item threshold parameters, it now is applied to the absolute item threshold parameters.
A bug occuring in pcmodel() when null categories are present and nullcats = "keep" was fixed. (Thanks to Oliver Prosperi for reporting this.)
The processing of the minimal category zero in the function rsmodel() was changed. Only if for all items, the minimal category is above zero, downcoding takes place. Otherwise, the missing minimal categories are treated as not observed, i.e., with a frequency of zero.
Major update with new model fitting functions (partial credit and rating scale model) and improved infrastructure for conditional maximum likelihood estimation (C implementation of elementary symmetric functions).
Partial credit models (PCMs) can be fitted with the function PCModel.fit(). The interface and return value is similar to that of RaschModel.fit().
Rating scale models (RSMs) can be fitted with the function RSModel.fit(). The interface and return value is similar to that of RaschModel.fit() and PCModel.fit().
The function elementary_symmetric_functions() for computing ESFs is extended and now part of the exported user interface. The R implementation for binary items up to order 2 is complemented by a C implementation for both binary and polytomous items up to order 1.
Due to numerical instabilities in the coefficients and standard errors between different architectures, the optimization method for Rasch/RSModel/PCModel.fit() was changed from nlm(...) to optim(..., method = "BFGS"). Consequently, the arguments reltol and maxit are used now instead of gradtol and iterlim. For backward compatibility RaschModel.fit() still supports the old arguments but might cease to do so in future releases.
YouthGratitude data from Froh, Fan, Emmons, Bono, Huebner, Watkins (2011, Psychological Assessment), provided by Jeff Froh and Jinyan Fan. Some approximate replication code is provided in the examples (the parts depending on lavaan are in \dontrun).elementary_symmetric_functions(). (An extended C implementation is under development and will be included in future releases.)Support of non-integer weights in btReg.fit(). To facilitate this, summary.paircomp() gained a weights argument so that optionally the weights are aggregated instead of observations counted.
Actually pass on nlm() arguments from RaschModel.fit(). Also support iterlim = 0, i.e., set up model at pre-specified parameters.
Added StereotypeThreat data from Wicherts, Conor, Hessen (2005, Journal of Personality and Social Psychology), provided by Jelte M. Wicherts. Replication code is provided in the examples (the parts depending on lavaan are in \dontrun).
New psychotools package containing all “base” infrastructure previously contained in psychotree. This is in order to provide both methods and data that can be reused by psychotree and the new package psychomix (as well as potentially further packages).
Classes: paircomp and associated methods.
Models: btReg.fit() and RaschModel.fit() and associated methods.
Data: Firstnames, GermanParties2009, Soundquality (previoulsy in psychotree) and VerbalAggression (new data, contained in other formatting in difR as verbal and lme4 as VerbAgg).