Changes in Version 0.7-1
o 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").
o Fix logicals of length > 1 in some if() constructs.
o Accept .replicates for user-provided model specifications in mptmodel().
Changes in Version 0.7-0
o 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.
o 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).
o 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.
o Added the "proCNI" multinomial processing tree specification in mptspec() for
mptmodel(). This provides the CNI model of moral dilemma judgment for
proscriptive norms.
Changes in Version 0.6-1
o New 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.
Changes in Version 0.6-0
o 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.
o Changed method to invert hessian for raschmodel(), rsmodel(), and pcmodel()
from qr.solve() to chol2inv(chol()).
o Added two demos on how to conduct simulation studies for score-based test
of measurement invariance.
o All IRT models now have a function to simulate IRT data, see rrm(), rrsm(),
rpcm(), rpl(), and rgpcm().
o 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
Changes in Version 0.5-1
o Added "IGNORE_RDIFF" flags in some examples in order to avoid showing
diffs due to small numeric deviations in some checks (especially on CRAN).
Changes in Version 0.5-0
o 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.
o The new gpcmodel() function interfaces mirt (see above) and fits (generalized)
partial credit models (GPCMs) by MML.
o The new plmodel() function interfaces mirt (see above) and fits various
parametric IRT logistic models using MML: 1PL (Rasch), 2PL, 3PL, 3PLu, and
4PL.
o New functions and eponymous classes guesspar(), and upperpar() to
extract/represent so-called guessing parameters and upper asymptote parameters
of IRT models.
o 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.
o Added new data set "ConspiracistBeliefs2016" from the Open Source Psychometrics
Project (2016).
o Added new simulated data set "Sim3PL" for fitting dichotomous IRT models,
especially the 3PL and 3PLu.
Changes in Version 0.4-3
o Conditionally register all estfun() and bread() S3 methods for model
objects, provided that the "sandwich" package is attached.
o Added native routine registration for esf.c.
o Use R version of elementary_symmetric_functions() by default on Win/i386
due to small numeric differences on that platform.
o 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.
o 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.
Changes in Version 0.4-2
o Added new data set "PairClustering" from Klauer (2006).
o Fixed replication code in example of StereotypeThreat (reported by
Ed Merkle).
o Basil Abou El-Komboz changed his name to Basil Komboz.
Changes in Version 0.4-1
o Properly imported grDevices and utils in NAMESPACE.
o 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.
Changes in Version 0.4-0
o 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.
o Bug fix in itempar() method for "raschmodel" objects if alias = FALSE.
In the previous version the methods had an erroneous trailing NA.
o Improved item names labeling in plot() method for "itemresp" objects
to conform with regionplot() function for IRT models.
o mscale<-() method for "itemresp" has been improved so that categories
can be easily collapsed (e.g., dichotomized).
Changes in Version 0.3-0
o 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.
o 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).
o 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.
o The worth() generic now internally calls the methods for itempar().
o 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".
o 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.
o 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.
o 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.
o 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.
o A bug occuring in pcmodel() when null categories are present and
nullcats = "keep" was fixed. (Thanks to Oliver Prosperi for reporting
this.)
o 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.
Changes in Version 0.2-0
o 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).
o Partial credit models (PCMs) can be fitted with the function
PCModel.fit(). The interface and return value is similar to that
of RaschModel.fit().
o 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().
o 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.
o 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.
Changes in Version 0.1-4
o Added YouthGratitude data from Froh, Fan, Emmons, Bono, Huebner, Watkins
(2011, PA), provided by Jeff Froh and Jinyan Fan. Some approximate
replication code is provided in the examples (the parts depending on
lavaan are in \dontrun).
Changes in Version 0.1-3
o Fully exported elementary_symmetric_functions(). (An extended C
implementation is under development and will be included in
future releases.)
Changes in Version 0.1-2
o 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.
o Actually pass on nlm() arguments from RaschModel.fit(). Also
support iterlim = 0, i.e., set up model at pre-specified parameters.
o Added StereotypeThreat data from Wicherts, Conor, Hessen (2005, JPSP),
provided by Jelte M. Wicherts. Replication code is provided in the
examples (the parts depending on lavaan are in \dontrun).
Changes in Version 0.1-1
o 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).
o Classes: "paircomp" and associated methods.
o Models: btReg.fit() and RaschModel.fit() and associated methods.
o Data: Firstnames, GermanParties2009, Soundquality (previoulsy in
psychotree) and VerbalAggression (new data, contained in other
formatting in difR/verbal and lme4/VerbAgg).